• See book's home page for info and buy links.
• Go to web version's table of contents.
4. Enjoying Your Mind
We’re fortunate enough to be able to observe minds in action all day long. You can make a fair amount of progress on researching the mind by paying close attention to what passes through your head as you carry out your usual activities.
Doing book and journal research on the mind is another story. Everybody has their own opinion—and everybody disagrees. Surprising as it may seem, at this point in mankind’s intellectual history, we have no commonly accepted theory about the workings of our minds.
In such uncharted waters, it’s the trip that counts, not the destination. At the very least, I’ve had fun pondering this chapter, in that it’s made me more consciously appreciative of my mental life. I hope it’ll do the same for you—whether or not you end up agreeing with me.
As in the earlier chapters, I’ll speak of a hierarchy of computational processes. Running from the lowest to the highest, I’ll distinguish among the following eight levels of mind, devoting a section to each level, always looking for connections to notions of computation.
• 4.1: Sensational Emotions. Sensation, action, and emotion.
• 4.2: The Network Within. Reflexes and learning.
• 4.3: Thoughts as Gliders and Scrolls. Thoughts and moods.
• 4.4: “I Am.” Self-awareness.
• 4.5: The Lifebox. Memory and personality.
• 4.6: The Mind Recipe. Cogitation and creativity.
• 4.7: What Do You Want? Free will.
• 4.8: Quantum Soul. Enlightenment.
One of my special goals here will be to make good on the dialectic triad embodied in the title, The Lifebox, the Seashell, and the Soul. As I mentioned earlier, there’s a tension between the computational “lifebox” view of the mind vs. one’s innate feeling of being a creative being with a soul. I want to present the possibility that the creative mind might be a kind of class four computation akin to a scroll-generating or cone-shell-pattern-generating cellular automaton.
But class four computation may not be the whole story. In the final section of the chapter I’ll discuss my friend Nick Herbert’s not-quite-relevant but too-good-to-pass-up notion of viewing the gap between deterministic lifebox and unpredictable soul as relating to a gap between the decoherent pure states and coherent superposed states of quantum mechanics.
• See book's home page for info and buy links.
• Go to web version's table of contents.
4.1: Sensational Emotions
Mind level one—sensation, action, and emotion—begins with the fact that, in the actual world, minds live in physical bodies. Having a body situated in a real world allows a mind to receive sensations from the world, to act upon the world, and to observe the effects of these actions.
We commonly use the term sensors for an organism’s channels for getting input information about the world, and the term effectors for the output channels through which the organism does things in the world—moving, biting, growing, and so on (see Figure 79).
Figure 79 : A Mind Connects Sensors to Effectors
The lump in between the sensors and the effectors represents the organism’s mind. The not-so-familiar word proprioception refers to the awareness of the relative positions of one’s joints and limbs.
Being embedded in the world is a given for any material entity—human, dog, robot, or boulder. Paradoxically enough, the goal of high-level meditative practices is to savor fully this basic level of being. Meditation often begins by focusing on the simplest possible action that a human takes, which is breathing in and out.
Although being embedded in the world comes naturally for physical beings, if you’re running an artificial-life (a-life) program with artificial organisms, your virtual creatures won’t be embedded in a world unless you code up some sensor and effector methods they can use to interact with their simulated environment.
The fact that something which is automatic for real organisms requires special effort for simulated organisms may be one of the fundamental reasons why artificial-life and simulated evolution experiments haven’t yet proved as effective as we’d like them to be. It may be that efforts to construct artificially intelligent programs for personal computers (PCs) will founder until the PCs are put inside robot bodies with sensors and effectors that tie them to the physical world.
In most of this chapter, I’m going to talk about the mind as if it resides exclusively in the brain. But really your whole body participates in your thinking.
Regarding the mind-brain-body distinction, I find it amusing to remember how the ancient Egyptians prepared a body for mummification (see Table 9). Before curing and wrapping the corpse’s muscles and bones, they’d eviscerate the body and save some key organs in so-called canopic jars. These jars were carved from alabaster stone, each with a fancy little sculpture atop its lid. Being a muscle, the heart was left intact within the body. And what about the brain? The ancient Egyptians regarded the brain as a worthless agglomeration of mucus; they pulled it out through the nose with a hooked stick!
Preserved in falcon-headed canopic jar
Preserved in jackal-headed canopic jar
Preserved in human-headed canopic jar
Preserved in baboon-headed canopic jar
Mummified with the body
Pulled out the nose and fed to the dogs
Table 9: What Ancient Egyptians Did with a Mummy’s Organs
The degree to which we’ve swung in the other direction can be seen by considering the science-fictional notion of preserving a person as a brain in a jar. If you can’t afford to court immortality by having your entire body frozen, the cryonicists at the Alcor Life Extension Foundation are glad to offer you what they delicately term the neuropreservation option—that is, they’ll freeze your brain until such a time as the brain-in-a-jar or grow-yourself-a-newly-cloned-body technology comes on line. But the underlying impulse is, of course, similar to that of the ancient Egyptians.
Anyway, my point was that in some sense you think with your body as well as with your mind. This becomes quite clear when we analyze the nature of emotions. Before taking any physical action, we get ourselves into a state of readiness for the given motions. The whole body is aroused. The brain scientist Rodolfo Llinas argues that this “premotor” state of body activation is what we mean by an emotion.77
One of the results of becoming socialized and civilized is that we don’t in fact carry out every physical action that occurs to us—if I did, my perpetually malfunctioning home and office PCs would be utterly thrashed, with monitors smashed in, wires ripped out, and beige cases kicked in. We know how to experience an emotional premotor state without carrying out some grossly inappropriate physical action. But even the most tightly controlled emotions manifest themselves in the face and voice box, in the lower back, in the heart rate, and in the secretions of the body’s glands.
One of the most stressful things you can do to yourself is to spend the day repeating a tape loop of thoughts that involves some kind of negative emotion. Every pulse of angry emotion puts your body under additional activation. But if that’s the kind of day you’re having—oh well, accept it. Being unhappy’s bad enough without feeling guilty about being unhappy!
Given the deep involvement of emotions with the body, it’s perhaps questionable whether a disembodied program can truly have an emotion—despite the classic example of HAL in 2001: A Space Odyssey. But, come to think of it, HAL’s emotionality was connected with his hardware, so in that sense he did have a body. Certainly humanoid robots will have bodies, and it seems reasonable to think of them as having emotions. Just as a person does, a robot would tend to power up its muscle motors before taking a step. Getting ready to fight someone would involve one kind of preparation, whereas preparing for a nourishing session of plugging into an electrical outlet would involve something else. It might be that robots could begin to view their internal somatic preparations as emotions. And spending too much time a prefight state could cause premature wear in, say, a robot’s lower-back motors.
A slightly different way to imagine robot emotions might be to suppose that a robot continually computes a happiness value that reflects its status according to various measures. In planning what to do, a robot might internally model various possible sequences of actions and evaluate the internal happiness values associated with the different outcomes—not unlike a person thinking ahead to get an idea of the emotional consequences of a proposed action. The measures used for a robot “happiness” might be a charged battery, no hardware error-messages, the successful completion of some tasks, and the proximity of a partner robot. Is human happiness so different?
• See book's home page for info and buy links.
• Go to web version's table of contents.
4.2: The Network Within
Looking back at Figure 79, what goes inside the lump between the sensors and the effectors? A universal automatist would expect to find computations in there. In this section, I’ll discuss some popular computational models of how we map patterns of sensory stimulation into patterns of effector activation.
Perhaps the simplest possible mind component is a reflex: If the inputs are in such and such a pattern, then take this or that action. The most rudimentary reflex connects each sensor input to an effector output. A classic example of this is a self-steering toy car, as shown in Figure 80.
Figure 80: A toy Car That Straddles a Stripe
The eyes stare down at the ground. When the right eye sees the stripe, this means the car is too far to the left and the car swerves right. Likewise for the other side.
Our minds include lots of simple reflexes—if something sticks in your throat, you cough. If a bug flies toward your eye, you blink. When your lungs are empty you breathe in; when they’re full you breathe out.
Microorganisms get by solely on reflexes—a protozoan swims forward until it bumps something, at which time it reverses direction. If the water holds the scent of food, the creature slows down, otherwise it swims at full speed. Bumping and backing, speeding and slowing down, the animalcule finds its way all around its watery environment. When combined with the unpredictable structures of the environment, a protozoan’s bump-and-reverse reflex together with its slow-down-near-food reflex are enough to make the creature seem to move in an individualistic way.
Looking through a microscope, you may even find yourself rooting for this or that particular microorganism—but remember that we humans are great ones for projecting personality onto the moving things we see. As shown in Figure 81, the internal sensor-to-effector circuitry of a protozoan can be exceedingly simple.
Figure 81: A Flagellate’s Swimming Reflexes
I’m drawing the little guy with a corkscrew flagellum tail that the can twirl in one direction or the other, at a greater or lesser speed. On sensor controls speed and the other sensor controls the direction of the twirling.
In the early 1980s, brain researcher Valentino Braitenberg had the notion of describing a series of thought experiments involving the motions of simple “vehicles” set into a plane with a few beacon lights.78 Consider for instance a vehicle with the reflex “approach any light in front of you until it gets too bright, then veer to the dimmer side.” If we were to set down a vehicle of this kind in a plane with three lights, we would see an interesting circling motion, possibly chaotic. Since Braitenberg’s time, many programmers have amused themselves either by building actual Braitenberg vehicles or by writing programs to simulate little worlds with Braitenberg vehicles moving around. If the virtual vehicles are instructed to leave trails on the computer screen, knotted pathways arise—not unlike the paths that pedestrians trample into new fallen snow. Figure 82 shows three examples of Braitenberg simulations.
Figure 82: Gnarly Braitenberg Vehicles
The artist and UCLA professor Casey Reas produced these images by simulating large numbers of Braitenberg vehicles that steer toward (or away from) a few “light beacons” placed in their plane worlds. These particular images are called Tissue, Hairy Red, and Path, and are based on the motions of, respectively, twenty-four thousand, two hundred thousand, and six hundred vehicles reacting to, respectively, three, eight, and three beacons. To enhance his images, Reas arbitrarily moves his beacons now and then. He uses his programs to create fine art prints of quite remarkable beauty (see www.groupc.net).
Although the Braitenberg patterns are beautiful, if you watch one particular vehicle for a period of time, you’ll normally see it repeating its behavior. In other words, these vehicles, each governed by a single reflex force, perform class two computations. What can we do to make their behavior more interesting?
I have often taught a course on programming computer games, in which students create their own interactive three-dimensional games.79 We speak of the virtual agents in the game as “critters.” A recurrent issue is how to endow a critter with interesting behavior. Although game designers speak of equipping their critters with artificial intelligence, in practice this can mean something quite simple. In order for a critter to have interesting and intelligent-looking behavior, it can be enough to attach two or, better, three reflex-driven forces to it.
Think back to our example in section 3.3: Surfing Your Moods of a magnet pendulum bob suspended above two or more repelling magnets. Three forces: gravity pulls the bob toward the center, and the two magnets on the ground push it away. The result is a chaotic trail.
But now suppose you want more than chaos from your critter. Suppose you want it to actually be good at doing something. In this case the simplest reflexes might not be enough. The next step above the reflex is to allow the use of so-called logic gates that take one or more input signals and combine them in various ways. If we allow ourselves to chain together the output of one gate as the input for another gate, we get a logical circuit. It’s possible to create a logical circuit to realize essentially any way of mapping inputs into outputs. We represent a simple “walking” circuit in Figure 83.
Figure 83: Logic Gates for a Walker
The goblet-shaped gates are AND gates, which output a value of True at the bottom if and only both inputs at the top are True. The little circles separating two of the gate input lines from the gates proper are themselves NOT gates. A NOT gate returns a True if and only if its input is False. If you set the walker down with one leg bent and one leg straight, it will continue alternating the leg positions indefinitely.
And of course we can make the circuits more and more complicated, eventually putting in something like an entire PC or more.
But a logic circuit may not be the best way to model a mind.
The actual computing elements of our brains are nerve cells, or neurons. A brain neuron has inputs called dendrites and a branching output line called an axon, as illustrated in Figure 84. Note, by the way, that an axon can be up to a meter long. This means that the brain’s computation can be architecturally quite intricate. Nevertheless, later in this chapter we’ll find it useful to look at qualitative CA models of the brain in which the axons are short and the neurons are just connected to their closest spatial neighbors.
Figure 84: A Brain Neuron
There are about a hundred billion neurons in a human brain. A neuron’s number of “input” dendrites ranges from several hundred to the tens of thousands. An output axon branches out to affect a few dozen or even a few hundred other cells.
In accordance with the sensor-mind-effector model, the neuron dendrites can get inputs from sensor cells as well as from other neurons, and the output axons can activate muscle cells as well as other neurons, as indicated in Figure 85.
Normally a nerve cell sends an all-or-nothing activation signal out along its axon. Whether a given neuron “fires” at a given time has to do with its activation level, which in turn depends upon the inputs the neuron receives through its dendrites. Typically a neuron has a certain threshold value, and it fires when its activation becomes greater than the threshold level. After a nerve cell has fired, its activation drops to a resting level, and it takes perhaps a hundredth of a second of fresh stimulation before the cell can become activated enough to fire again.
The gap where axons meet a dendrite is a synapse. When the signal gets to a synapse it activates a transmitter chemical to hop across the gap and stimulate the dendrites of the receptor neuron. The intensity of the stimulation depends on the size of the bulb at the end of the axon and on the distance across the synapse. To make things a bit more complicated, there are in fact two kinds of synapse: excitatory and inhibitory. Stimulating an inhibitory synapse lowers the receptor cell’s activation level.
Most research on the human brain focuses on the neocortex, which is the surface layer of the brain. This layer is about an eighth of an inch thick. If you could flatten out this extremely convoluted surface, you’d get something like two square feet of neurons—one square foot per brain-half.
Figure 85: Brain Neurons Connected to Sensors and Effectors.
Our big flat neocortical sheets can be regarded as a series of layers. The outermost layer is pretty much covered with a tangle of dendrites and axons—kind of like the zillion etched connector lines you see on the underside of a circuit board. Beneath this superficial layer of circuitry, one can single out three main zones: the upper layers, the middle layers, and the deep layers, as illustrated in Figure 86.
Roughly speaking, sensory inputs arrive at the middle layer, which sends signals to the upper layer. The upper layer has many interconnections among its own neurons, but some of its outputs also trickle down to the deep layer. And the neurons in the deep layer connect to the body’s muscles. This is indicated in Figure 86, but we make the flow even clearer in Figure 87.
As usual in biology, the actual behavior of a living organic brain is funkier and gnarlier and much more complicated than any brief sketch. Instead of there being one kind of brain neuron, there are dozens. Instead of there being one kind of synapse transmission chemical, there are scores of them, all mutually interacting. Some synapses operate electrically rather than chemically, and thus respond much faster. Rather than sending a digital pulse signal, some neurons may instead send a graded continuous-valued signal. In addition, the sensor inputs don’t really go directly to the neocortex; they go instead to an intermediate brain region called the thalamus. And our throughput diagram is further complicated by the fact that the upper layers feed some outputs back into the middle layers and the deep layers feed some outputs back into the thalamus. Furthermore, there are any number of additional interactions among other brain regions; in particular, the so-called basal ganglia have input-output loops involving the deep layers. But for our present purposes the diagrams I’ve given are enough.
Figure 86: The Neocortex, Divided into Upper, Middle and Deep Layers.
A peculiarity of the brain-body architecture is that the left half of the brain handles input and output for the right side of the body, and vice versa.
When trying to create brainlike systems in their electronic machines, computer scientists often work with a switching element also called a neuron—which is of course intended to model a biological neuron. These computer neurons are hooked together into circuits known as neural nets or neural networks.
As drawn in Figure 88, we think of the computer neuron as having some number of input lines with a particular weight attached to each input. The weights are continuous real numbers ranging from negative one to positive one. We think of negatively weighted inputs as inhibiting our artificial neurons and positive values as activating them. At any time, the neuron has an activation level that’s obtained by summing up the input values times the weights. In addition, each neuron has a characteristic threshold value Th. The neuron’s output depends on how large its activation level is relative to its threshold. This is analogous to the fact that a brain neuron fires an output along its axon whenever its activation is higher than a certain value.
Figure 87: A Simplified Diagram of the Neocortical Layers’ Throughput
U, M, and D are, respectively, the upper, middle, and deep layers of the neocortex.
Computer scientists have actually worked with two different kinds of computer neuron models. In the simple version, the neuron outputs are only allowed to take on the discrete values zero and one. In the more complex version, the neuron outputs can take on any real value between zero and one. Our figure indicates how we might compute the output values from the weighted sum of the inputs in each case.
Most computer applications of neural nets use neurons with continuous-valued outputs. On the one hand this model is more removed from the generally discrete-valued-output behavior of biological neurons—given that a typical brain neuron either fires or doesn’t fire, its output is more like a single-bit zero or one. On the other hand, our computerized neural nets use so many fewer neurons than does the brain that it makes sense to get more information out of each neuron by using the continuous-valued outputs. In practice, it’s easier to tailor a smallish network of real-valued-output neurons to perform a given task.80
One feature of biological neurons that we’ll ignore for now is the fact that biological neurons get tired. That is, there is a short time immediately after firing when a biological neuron won’t fire again, no matter how great the incoming stimulation. We’ll come back to this property of biological neurons in the following section 4.3: Thoughts as Gliders and Scrolls.
In analogy to the brain, our artificial neural nets are often arranged in layers. Sensor inputs feed into a first layer of neurons, which can in turn feed into more layers of neurons, eventually feeding out to effectors. The layers in between the sensors and the final effector-linked layers are sometimes called hidden layers. In the oversimplified model of the neocortex that I drew in Figures 86 and 87, the so-called upper layer would serve as the hidden layer.
What makes neural nets especially useful is that it’s possible to systematically tweak the component neurons’ weights in order to make the system learn certain kinds of task.
Figure 88: Computer Neuron Models
A neuron with a threshold value computes the weighted sum of its inputs and uses this sum to calculate its output. In this figure, the threshold is Th, and the activation level Ac is the weighted sum of the inputs is w1•i1 + w2•i2 + w3•i3. Below the picture of the neuron we graph two possible neuron response patterns, with the horizontal axes representing the activation level Ac and the vertical axes representing the output value o. In the upper “discrete” response pattern, the output is one if Ac is greater than Th, and zero otherwise. In the lower “continuous” response pattern, the output is graded from zero to one, depending on where Ac stands relative to Th.
In order to get an idea of how neural nets work, consider a face-recognition problem that I often have my students work on when I teach a course on artificial intelligence. In this experiment, which was devised by the computer scientist Tom Mitchell of Carnegie Mellon University, we expose a neural network to several hundred small monochrome pictures of people’s faces.81 To simplify things, the images are presented as easily readable computer files that list a grayscale value for each pixel. The goal of the experiment, as illustrated in Figure 89, is for the network to recognize which faces are smiling, which are frowning, and which are neutral.
As it turns out, there’s a systematic method for adjusting the weights and thresholds of the neurons. The idea is that you train the network on a series of sample inputs for which the desired target output is known. Suppose you show it, say, a test face that is to be known to be smiling. And suppose that the outputs incorrectly claim that the face is frowning or neutral. What you then do is to compute the error term of the outputs. How do I define the error? We’ll suppose that if a face is frowning, I want an output of 0.9 from the “frown” effector, and outputs of 0.1 from the other two—and analogous sets of output for the neutral and smiling cases. (Due to the S-shaped nature of the continuous-valued-output neurons’ response curve, it’s not practical to expect them to actually produce exact 0.0 or 1.0 values.)
Given the error value of the network on a given face, you can work your way backward through the system, using a so-called back-propagation algorithm to determine how best to alter the weights and thresholds on the neurons in order to reduce the error. Suppose, for instance, the face was frowning, but the smile output was large. You look at the inputs and weights in the smile neuron to see how its output got so big. Perhaps one of the smile neuron inputs has a high value, and suppose that the smile neuron also has a high positive weight on that input. You take two steps to correct things. First of all, you lower the smile neuron’s weight for that large-valued input line, and second, you go back to the hidden neuron that delivered that large value to the smile neuron and analyze how that hidden neuron happened to output such a large value. You go over its input lines and reduce the weights on the larger inputs. In order for the process of back propagation to work, you do it somewhat gently, not pushing the weights or thresholds too far at any one cycle. And you iterate the cycle many times.
So you might teach your neural net to recognize smiles and frowns as follows. Initialize your network with some arbitrary small weights and thresholds, and train it with repeated runs over a set of a hundred faces you’ve already determined to be smiling, frowning, or neither. After each face your back-propagation method computes the error in the output neurons and readjusts the network weights that are in some sense responsible for the error. You keep repeating the process until you get to a run where the errors are what you consider to be sufficiently small—for the example given, about eighty runs might be enough. At this point you’ve finished training your neural net. You can save the three thousand or so trained weights and thresholds into a file.
You may be reminded of the discussion in section 3.6: How We Got So Smart of search methods for finding peaks in a fitness landscape. The situation here is indeed similar. The “face weight” landscape example ranges over all the possible sets of neuron weights and thresholds. For each set of three thousand or so parameters, the fitness value measures how close the network comes to being a 100 percent right in discerning the expressions of the test faces. Obviously it’s not feasible to carry out an exhaustive search of this enormous parameter space, so we turn to a search method—in this case a method called back-propagation. As it turns out, back-propagation is in fact a special hill-climbing method that, for any given set of parameters, finds the best nearby set of parameters in the neural net’s fitness landscape.
Figure 89: A Neural Net to Recognize Smiles and Frowns
We’ll use continuous-valued-output neurons, so at the right-hand side our net returns a real-number “likelihood” value for each of the three possible facial expressions being distinguished. The inputs on the left consist of rasterized images of faces, 30 x 32 pixels in size, with each of the 960 pixels containing a grayscale value ranging between zero and one. In the middle we have a “hidden layer” of three neurons. Each neuron takes input from each of the 960 pixels and has a single output line. On the right we have three effector neurons (drawn as three faces), each of which takes as input the outputs of the three hidden-layer neurons. For a given input picture, the network’s “answer” regarding that picture’s expression corresponds to the effector neuron that returns the greatest value. The total number of weights in this system is 3^960 + 33, or 2,889, and if we add in the threshold values for our six neurons, we get 2,895. (By the way, there’s no particular reason why I have three hidden neurons. The network might do all right with two or four hidden neurons instead.)
Now comes the question of how well your net’s training can generalize. Of course it will score fine on those hundred faces that it looked at eighty times each—but what happens when you give it new, previously unseen pictures of faces? If the faces aren’t radically different from the kinds of faces the net was trained upon, the net is likely to do quite well. The network has learned something, and the knowledge takes the form of some three thousand real numbers.
Figure 90: Generalized Face Recognizer
The first box recognizes the face’s position. The position information is fed with the face information to the three specialized expression recognizers, and the appropriate one responds.
But if, for instance, all the training faces were seen full-on, the network isn’t going to do well on profile faces. Or if all the training faces were of people without glasses, the network probably won’t be good at recognizing the facial expressions of glasses-wearers.
What to do? One approach is to build up deeper layers of neural networks. We might, for instance, use a preliminary network to decide which way a face is looking, sorting them into left-profile, full, and right-profile faces. And then we could train a separate expression-recognizing network for each of the three kinds of face positions. With a little cleverness with the wiring, the whole thing could be boxed up as a single multilayer neural net, as suggested by Figure 90.
Note here that it can actually be quite a job to figure out how best to break down a problem into neural-net-solvable chunks, and to then wire the solutions together. Abstractly speaking, one could simply throw a very general neural net at any problem—for instance, you might give the arbitrarily-positioned-face-recognition problem a neural net with perhaps two hidden layers of neurons, also allowing input lines to run from all the pixels into the neurons of the second hidden layer, and hope that with enough training and back-propagation the network will eventually converge on a solution that works as well as using the first layer to decide on the facial orientation and using the second layer’s neurons to classify the expression of each facial orientation. Training a net without a preconceived design is feasible, but it’s likelier to take longer than using some preliminary analysis and assembling it as described above.
Custom-designed neural nets are widely used—the U.S. Post Office, for instance, uses a neural net program to recognize handwritten ZIP codes. But the designers did have to put in quite a bit of thought about the neural net architecture—that is, how many hidden layers to use, and how many neurons to put in each of the hidden layers.
In the 2010s, software designers began using more than one hidden layer in their neural nets, as well as tailoring the connections for the targeted application. They’ve also begun using nonlinear response curves, that is, the effect of an input value might vary, say, with the square of the value, instead of with a simple multiple of the value. This arsenal of new techniques falls under the rubric of “deep learning.” Impressive strides forward are being made.
Computer scientists like to imagine building programs or robots that can grow up and learn and figure things out without any kind of guiding input. In so-called unsupervised learning, there’s no answer sheet to consult. If, for instance, you learn to play Ping-Pong simply by playing games against an opponent, your feedback will consist of noticing which of your shots goes bad and which points you lose. Nobody is telling you things like, “You should have tilted the paddle a little to the right and aimed more toward the other side’s left corner.” And, to make the situation even trickier, it may be quite some time until you encounter a given situation again. Tuning your neural net with unsupervised learning is a much harder search problem, and specifying the search strategy is an important part of the learning program design—typical approaches might include hill-climbing and genetic programming.
There’s also a metasearch issue of trying out various neural net architectures and seeing which works best. But as your problem domains get more sophisticated, the fitness evaluations get more time-consuming, particularly in the unsupervised learning environments where you essentially have to play out a whole scenario in order to see how well your agent will do. The search times for effective learning can be prohibitively long.
It’s worth noting here that a human or an animal is born with nearly all of the brain’s neural net architecture already in place. It’s not as if each of us has to individually figure out how to divide the brain’s hundred billion neurons into layers, how to parcel these layers into intercommunicating columns, and how to connect the layers to the thalamus, the sense organs, the spine, the basal ganglia, etc. The exceedingly time-consuming searches over the space of possible neural architectures is something that’s happened over millions of years of evolution—and we’re fortunate enough to inherit the results.
The somewhat surprising thing is how often a neural net can solve what had seemed to be a difficult AI problem. Workers in the artificial-intelligence field sometimes say, “Neural nets are the second-best solution to any problem.”82
Given a reasonable neural net architecture and a nice big set of training examples, you can teach a neural net to solve just about any kind of problem that involves recognizing standard kinds of situations. And if you have quite a large amount of time, you can even train neural nets to carry out less clearly specified problems, such as, for instance, teaching a robot how to walk on two legs. As with any automated search procedure, the neural net solutions emerge without a great deal of formal analysis or deep thought.
The seemingly “second-best” quality of the solution has to do with the feeling that a deep learning neural net solution is somewhat clunky, ad hoc, and brute-force. It’s not as if the designer has come up with an elegant, simple algorithm based upon a fundamental understanding of the problem in question. The great mound of network weights has an incomprehensible feel to it.
It could be that it’s time to abandon our scientific prejudices against complicated solutions. In the heroic nineteenth and twentieth centuries of science, the best solution of a problem often involved a dramatic act of fundamental understanding—one has only to think of the kinds of formulas that traditionally adorn the T-shirts of physics grad students: Maxwell’s equations for electromagnetism, Einstein’s laws for relativity, Schrodinger’s wave equation for quantum mechanics. In each case, we’re talking about a concise set of axioms from which one can, in a reasonable amount of time, logically derive the answers to interesting toy problems.
But the simple equations of physics don’t provide feasible solutions to many real-world problems—the laws of physics, for instance, don’t tell us when the big earthquake will hit San Jose, California, and it wouldn’t even help to know the exact location of all the rocks underground. Physical systems are computationally unpredictable. The laws provide, at best, a recipe for how the world might be computed in parallel particle by particle and region by region. But—unless you have access to some so-far-unimaginable kind of computing device that simulates reality faster than the world does itself—the only way to actually learn the results is to wait for the actual physical process to work itself out. There is a fundamental gap between T-shirt-physics-equations and the unpredictable and PC-unfeasible gnarl of daily life.
One of the curious things about neural nets is that our messy heap of weights arises from a rather simple deterministic procedure. Just for the record, let’s summarize the factors involved.
• The network architecture, that is, how many neurons we use, and how they’re connected to the sensor inputs, the effector outputs, and to one another.
• The specific implementation of the back-propagation algorithm to be used—there are numerous variants of this algorithm.
• The process used to set the arbitrary initial weights—typically we use a pseudorandomizer to spit out some diverse values, perhaps a simple little program like a CA. It’s worth noting that if we want to repeat our experiment, we can set the pseudorandomizer to the same initial state, obtaining the exact same initial weights and thence the same training process and the same eventual trained weights.
• The training samples. In the case of the expression-recognition program, this would again be a set of computer files of face images along with a specification as to which expression that face is considered to have.
In some sense the weights are summarizing the information about the sample examples in a compressed form—and compressed forms of information are often random-looking and incomprehensible. Of course it might be that the neural net’s weights would be messy even if the inputs were quite simple. As we’ve seen several times before, it’s not unusual for simple systems to generate messy-looking patterns all on their own—remember Wolfram’s pseudorandomizing cellular automaton Rule 30 and his glider-producing universally computing Rule 110.
We have every reason to suppose that, at least in some respects, the human brain functions similarly to a computer-modeled neural network. Although, as I mentioned earlier, much of the brain’s network comes preinstalled, we do learn things—the faces of those around us, the words of our native language, skills like touch-typing or bicycle-riding, and so on.
First of all, we might ask how a living brain goes about tweaking the weights of the synaptic connections between axons and dendrites. One possibility is that the changes are physical. A desirable synapse might be enhanced by having the axon’s terminal bulb grow a bit larger or move a bit closer to the receptor dendrite, and an undesirable synapse’s axon bulb might shrivel or move away. The virtue of physical changes is that they stay in place. But it’s also possible that the synapse tweaking is something subtler and more biochemical.
Second, we can ask how our brains go about deciding in which directions the synapse weights should be tweaked.
Do we use back-propagation, like a neural net? This isn’t so clear.
A very simple notion of human learning is called Hebbian learning, after Canadian neuropsychologist Donald O. Hebb, who published an influential book called The Organization of Behavior in 1949. This theory basically says that the more often a given synapse fires, the stronger its weight becomes. If you do something the right way over and over, that behavior gets “grooved in.” Practice makes perfect. It may be that when we mentally replay certain kinds of conversations, we’re trying to simulate doing something right.
This said, it may be that we do some back-propagation as well. In an unsupervised learning situation, such as when you are learning to play tennis, you note an error when, say, you hit the ball into the net. But you may not realize the error is precisely because you rotated your wrist too far forward as you hit. Rather than being able to instantly back-propagate the information to your wrist-controlling reflexes, you make some kind of guess about what you did wrong and back-propagate that.
More complicated skills—like how to get along with the people in your life—can take even longer to learn. For one thing, in social situations the feedback is rather slow. Do you back-propagate the error term only the next day when you find out that you goofed, at that time making your best guess as to which of your behaviors produced the bad results?
As well as using Hebbian learning or back-propagation, we continually carry out experiments, adjusting our neural weights on the fly. To an outsider, this kind of activity looks like play. When puppies romp and nip, they’re learning. If you go off and hit a ball against a wall a few hundred times in a row, you’re exploring which kind of stroke gives the best results. What children learn in school isn’t so much the stuff the teachers say as it is the results of acting various different ways around other people.
In terms of Wolfram’s fourfold classification, what kinds of overall computation take place as you adjust the weights of your brain’s neural networks?
Clear-cut tasks like learning the alphabet are class one computations. You repeat a deterministic process until it converges on a fixed point.
But in many cases the learning is never done. Particularly in social situations, new problems continue to arise. Your existing network weights need to be retuned over and over. Your network-tuning computation is, if you will, a lifelong education. The educational process will have aspects of class two, three, and four.
Your life education is class two, that is, periodic, to the extent that you lead a somewhat sheltered existence, perhaps by getting all your information from newspapers or, even more predictably, from television. In this kind of life, you’re continually encountering the same kinds of problems and solving them in the same kinds of ways.
If, on the other hand, you seek out a wider, more arbitrary range of different inputs, then your ongoing education is more of a class three process.
And, to the extent that you guide yourself guided along systematic yet gnarly paths, you’re carrying out a class four exploration of knowledge. Note that in this last case, it may well be that you’re unable to consciously formulate the criteria by which you guide yourself—indeed, if your search results from an unpredictable class four computation, this is precisely the case.
The engaged intellectual life is a never-ending journey into the unknown. Although the rules of our neuronal structures are limited in number, the long-term consequences of the rules need not be boring or predictable.
• See book's home page for info and buy links.
• Go to web version's table of contents.
4.3: Thoughts as Gliders and Scrolls
So far we’ve only talked about situations in which a brain tries to adjust itself so as to deal with some external situation. But the life of the mind is much more dynamic. Even without external input, the mind’s evolving computations are intricate and unpredictable.
Your brain doesn’t go dark if you close your eyes and lie down in a quiet spot. The thoughts you have while driving your car don’t have much to do with the sensor-effector exigencies of finding your way and avoiding the other vehicles.
A little introspection reveals that we have several different styles of thought. In particular, I’d like to distinguish between two modes that I’ll call trains of thought and thought loops.
By trains of thought, I mean the free-flowing and somewhat unpredictable chains of association that the mind produces when left on its own. Note that trains of thoughts need not be formulated in words. When I watch, for instance, a tree branch bobbing in the breeze, my mind plays with the positions of the leaves, following them and automatically making little predictions about their motions. And then the image of the branch might be replaced by a mental image of a tiny man tossed up high into the air. His parachute pops open and he floats down toward a city of lights. I recall the first time I flew into San Jose and how it reminded me of a great circuit board. I remind myself that I need to see about getting a new computer soon, and then in reaction I think about going for a bicycle ride. The image of the bicyclist on the back of the Rider brand of playing cards comes to mind, along with a thought of how Thomas Pynchon refers to this image in Gravity’s Rainbow. I recall the heft of Pynchon’s book, and then think of the weights that I used to lift in Louisville as a hopeful boy forty-five years ago, the feel and smell of the rusty oily weight bar still fresh in my mind.
By thought loops, I mean something quite different. A thought loop is some particular sequence of images and emotions that you repeatedly cycle through. Least pleasant—and all too common—are thought loops having to do with emotional upsets. If you have a disagreement with a colleague or a loved one, you may find yourself repeating the details of the argument, summarizing the pros and cons of your position, imagining various follow-up actions, and then circling right back to the detailed replay of the argument. Someone deep in the throes of an argument-provoked thought loop may even move their lips and make little gestures to accompany the remembered words and the words they wish they’d said.
But many thought loops are good. You often use a thought loop to solve a problem or figure out a course of action, that is, you go over and over a certain sequence of thoughts, changing the loop a bit each time. And eventually you arrive at a pattern you like. Each section of this book, for instance, is the result of some thought loops that I carried around within myself for weeks, months, or even years before writing them down.
Understanding a new concept also involves a thought loop. You formulate the ideas as some kind of pattern within your neurons, and then you repeatedly activate the pattern until it takes on a familiar, grooved-in feel.
Viewing a thought loop as a circulating pattern of neuronal activation suggests something about how the brain might lay down long-term memories. Certainly a long-term memory doesn’t consist of a circulating thought loop that never stops. Long-term memory surely involves something more static. A natural idea would be that if you think about something for a little while, the circulating thought-loop stimulation causes physical alterations in the synapses that can persist even after the thought is gone. This harks back to the Hebbian learning I mentioned in the section 4.2: The Network Within, whereby the actual geometry of the axons and dendrites might change as a result of being repeatedly stimulated.
Thus we can get long-term memory from thought loops by supposing that a thought loop reinforces the neural pathway that it’s passing around. And when a relevant stimulus occurs at some later time, the grooved-in pathway is activated again.
Returning to the notion of free-running trains of thought, you might want to take a minute here and look into your mind. Watching trains of thought is entertaining, a pleasant way of doing nothing. But it’s not easy.
I find that what usually breaks off my enjoyment of my thought trains is that some particular thought will activate one of my thought loops to such an extent that I put all my attention into the loop. Or it may be that I encounter some new thought that I like so much that I want to savor it, and I create a fresh thought loop of repeating that particular thought to myself over and over. Here again my awareness of the passing thought trains is lost.
Figure 91: CA Made of Neurons
The dark, curvy eight-pointed stars are the neurons
This process is familiar for students of the various styles of meditation. In meditation, you’re trying to stay out of the thought loops and to let the trains of thought run on freely without any particular conscious attachment. Meditators are often advised to try to be in the moment, rather than in the past or the future—brooding over the past and worrying about the future are classic types of thought loops. One way to focus on the present is to pay close attention to your immediate body sensations, in particular to be aware of the inward and outward flow of your breath.
In understanding the distinction between trains of thought and thought loops, it’s useful to consider some computational models.
One of the nicest and simplest computer models of thought trains is a cellular automaton rule called Brian’s Brain. It was invented by the cellular automatist and all-around computer-fanatic Brian Silverman.
Brian’s Brain is based on a feature of the brain’s neurons that, for simplicity’s sake, isn’t normally incorporated into the computer-based neural nets described in section 4.2. This additional feature is that actual brain neurons can be tired, out of juice, too pooped to pop. If a brain neuron fires and sends a signal down its axon, then it’s not going to be doing anything but recuperating for the next tenth of a second or so.
With this notion in mind, Silverman formulated a two-dimensional CA model based on “nerve cells” that have three states: ready, firing, and resting. Rather than thinking of the cells as having distinct input and output lines, the CA model supposes that the component cells have simple two-way links; to further simplify things, we assume that each cell is connected only to its nearest neighbors, as indicated in Figure 91.
Figure 92: A Glider in the Brian’s Brain Rule
The Brian’s Brain rule updates all the cells in parallel, with a given cell’s update method depending on which state the cell happens to be in.
• Ready. This is the ground state where neurons spend most of their time. At each update, each cell counts how many (if any) of its eight immediate neighbors is in the firing state. If exactly two neighbors are firing, a ready cell switches to the firing state at the next update. In all other cases, a ready cell stays ready.
• Firing. This corresponds to the excited state where a neuron stimulates its neighbors. After being in a firing state, a cell always enters the resting state at the following update.
• Resting. This state follows the firing state. After one update spent in the resting state, a cell returns to the ready state.
The existence of the resting-cell state makes it very easy for moving patterns of activation to form in the Brian’s Brain rule. Suppose, for instance, that we have two light-gray firing cells backed up by two dark-gray resting cells. You might enjoy checking that the pattern will move in the direction of the firing cells, as indicated in Figure 92.
Brian’s Brain is rich in the moving patterns that cellular automatists call gliders. It’s a classic example of a class four computation and is nicely balanced between growth and death. Silverman himself once left a Brian’s Brain simulation running untended on an early Apple computer for a year—and it never died down. The larger gliders spawn off smaller gliders; occasionally gliders will stretch out a long thread of activation between them. As suggested by Figure 93, we find gliders moving in all four of the CA space’s directions, and there are in fact small patterns called butterflies that move diagonally as well.
I don’t think it’s unreasonable to suppose that the thought trains within your brain may result from a process somewhat similar to the endless flow of the gliding patterns in the Brian’s Brain CA. The trains of thought steam along, now and then colliding and sending off new thought trains in their wake.
Figure 93: The Brian’s Brain Cellular Automaton
The white, light gray, and dark gray cells are, respectively, in the ready, firing, and resting states. If the picture were animated, you would see the patterns moving horizontally and vertically, with the light gray edges leading the dark gray tails, and with new firing cells dying and being born.83
When some new sensation comes in from the outside, it’s like a cell-seeding cursor-click on a computer screen. A few neurons get turned on, and the patterns of activation and inhibition flow out from there.
A more complicated way to think of thought trains would be to compare the brain’s network of neurons to a continuous-valued CA that’s simulating wave motion, as we discussed back in section 2.2: Everywhere at Once. Under this view, the thought trains are like ripples, and new input is like a rock thrown into a pond.
What about recurrent thoughts—the topics that you obsess upon, the mental loops that you circle around over and over? Here another kind of CA rule comes to mind: Zhabotinsky-style scrolls. You may recall that we’ve already discussed these ubiquitous forms in both the context of morphogenesis as well as in connection with ecological simulations of population levels. I show yet another pair of these images in Figure 94.
Figure 94: More CA Scroll Patterns
These two images were generated using variations of continuous-valued activator-inhibitor rules suggested by, respectively, Arthur Winfree and Hans Meinhardt.
Scroll-type patterns often form when we have an interaction between activation and inhibition, which is a good fit for the computations of the brain’s neurons. And, although I haven’t yet mentioned the next fact, it’s also the case that scroll patterns most commonly form in systems where the individual cells can undergo a very drastic change from a high level of activation to a low level—which is also a good fit for neuronal behavior. A neuron’s activation levels rise to a threshold value, it fires, and its activation abruptly drops.
Recall that when CAs produce patterns like Turing stripes and Zhabotinsky scrolls, we have the activation and inhibition diffusing at different rates. I want to point out that the brain’s activation and inhibition signals may also spread at different rates. Even if all neural activation signals are sent down axons at the same rate, axons of different length take longer to transmit a signal. And remember that there’s a biochemical element to the transmission of signals across synapses, so the activator and inhibitor substances may spread and take effect at different rates.
Summing up, I see my thought patterns as being a combination of two types of processes: a discrete gliderlike flow of thought trains overlaid upon the smoother and more repetitive cycling of my thought loops. The images in Figure 95 capture something of what I have in mind.
Figure 95: Cellular Automaton Patterns Like a Mind with Thoughts and Obsessions
The first image shows the Brain Hodge rule, where discrete Brian’s Brain gliders cruise across a sea of Hodgepodge scrolls. The cells hold two activators, one for each rule, and the Brain activator stimulates the Hodgepodge activator. The second image shows the Boiling Cubic Wave rule. Here we have a nonlinear wave-simulating rule that makes ripples akin to thought loops. The nonlinearity of the wave is a value that varies from cell to cell and obeys a driven heat rule, producing a “boiling” effect like moving trains of thought layered on top of the wavy thought loops. As it so happens, the thought trains have the ability to bend around into square scrolls.
I’ll confess that neither of these images precisely models my conception of brain activity—I’d really prefer to see the gliders etching highways into the scrolls and to see the dense centers and intersections of the scrolls acting as seed points for fresh showers of gliders. But I hope the images give you a general notion of what I have in mind when I speak of thought trains moving across a background of thought loops.
I draw inspiration from the distinction between fleeting trains of thought and repeating thought loops. When I’m writing, I often have a fairly clear plan for the section I’m working on. As I mentioned above, this plan is a thought loop that I’ve been rehearsing for a period of time. But it sometimes happens that once I’m actually doing the writing, an unexpected train of thought comes plowing past. I treat such unexpected thoughts as gifts from the muse, and I always give them serious consideration.
I remember when I was starting out as a novelist, I read an advice-to-writers column where an established author said something like, “From time to time, you’ll be struck with a completely crazy idea for a twist in your story. A wild hair that totally disrupts what you had in mind. Go with it. If the story surprises you, it’ll surprise the reader, too.” I never forgot that advice.
This said, as a practical matter, I don’t really work in every single oddball thought I get, as at some point a work can lose its coherence. But many of the muse’s gifts can indeed be used.
Who or what is the muse? For now, let’s just say the muse is the unpredictable but deterministic evolution of thought trains from the various inputs that you happen to encounter day to day. The muse is a class four computation running in your brain.
People working in any kind of creative endeavor can hear the muse. You might be crafting a novel, an essay, a PowerPoint presentation, a painting, a business proposal, a computer program, a dinner menu, a decoration scheme, an investment strategy, or a travel plan. In each case, you begin by generating a thought loop that describes a fairly generic plan. And over time you expand and alter the thought loop. Where do the changes come from? Some of them are reached logically, as predictable class two computations closely related to the thought loop. But the more interesting changes occur to you as unexpected trains of thought. Your plan sends out showers of gliders that bounce around your neuronal space, eventually catching your attention with new configurations. Straight from the muse.
One seeming problem with comparing thoughts to moving patterns in cellular automata is that the space of brain neurons isn’t in fact structured like a CA. Recall that axons can be up to a meter long, and information flow along an axon is believed to be one-way rather than two-way.
But this isn’t all that big a problem. Yes, the connectivity of the brain neurons is more intricate than the connectivity of the cells in a CA. But our experiences with the universality of computational processes suggests that the same general kinds of patterns and processes that we find in CAs should also occur in the brain’s neural network. We can expect to find class one patterns that die out, class two patterns that repeat, chaotic class three patterns, rapidly moving gliderlike class four patterns, and the more slowly moving scroll-like class four patterns.
Visualize a three-dimensional CA running a CA rule rich in gliders and scrolls (see Figure 96). Think of the cells as nodes connected by strings to their neighbors. Now stretch a bunch of the strings and scramble things around, maybe drop the mess onto a tabletop, and shuffle it. Then paste the tangle into a two-foot-square mat that’s one-eighth of an inch thick. Your neocortex! The glider and scroll patterns are still moving around in the network, but due to the jumbled connections among the cells, a brain scan won’t readily reveal the moving patterns. The spatial arrangement of the neurons doesn’t match their connectivity. But perhaps some researchers can notice subtler evidences of the brain’s gliders and scrolls.
Figure 96: Another Three-Dimensional CA with Scrolls
Here’s a Winfree-style CA running on a three-dimensional array of cells. The surface patterns change rather rapidly, as buried scrolls boil up and move across it. Gnarly, dude!
At this point we’re getting rather close to the synthesizing “Seashell” element of The Lifebox, the Seashell, and the Soul. To be quite precise, I’m proposing that the brain is a CA-like computer and that the computational patterns called gliders and scrolls are the basis of our soulful mental sensations of, respectively, unpredictable trains of thought and repetitive thought-loops.
If this is true, does it make our mental lives less interesting? No. From whence, after all, could our thoughts come, if not from neuronal stimulation patterns? From higher-dimensional ectoplasm? From telepathic dark matter? From immortal winged souls hovering above the gross material plane? From heretofore undetected subtle energies? It’s easier to use your plain old brain.
Now, don’t forget that many or perhaps most complex computations are unpredictable. Yes, our brains might be carrying out computations, but that doesn’t mean they’ll ever cease surprising us.
I’m not always as happy as I’d like to be, and the months when I’ve been working on this chapter have been especially challenging. My joints and muscles pain me when I program or write, I had the flu for a solid month, my wife’s father has fallen mortally ill, my country’s mired in war, I’m anxious about finding a way to retire from the grind of teaching computer science, and frankly I’m kind of uptight about pulling off this rather ambitious book. I could complain for hours! I’m getting to be an old man.
Scientist that I am, I dream that a deeper understanding of the mind might improve my serenity. If I had a better model of how my mind works, maybe I could use my enhanced understanding to tweak myself into being happy. So now I’m going to see if thinking in terms of computation classes and chaotic attractors can give me any useful insights about my moods. I already said a bit about this in section 3.3: Surfing Your Moods, but I want to push it further.
Rather than calling the mind’s processes class four computations, we can also refer to them as being chaotic. Although a chaotic process is unpredictable in detail, one can learn the overall range of behaviors that the system will display. This range is, again, what we call a chaotic attractor.
A dynamic process like the flow of thought wanders around a state space of possibilities. To the extent that thought is a computation, the trajectory differs from random behavior in two ways. First of all, the transitions from one moment to the next are deterministic. And secondly, at any given time the process is constrained to a certain range of possibilities—the chaotic attractor. In the Brian’s Brain CA, for instance, the attractor is the behavior of having gliders moving about with a characteristic average distance between them. In a scroll, the attractor is the behavior of pulsing out rhythmic bands.
As I mentioned in section 2.4: The Meaning of Gnarl, a class one computation homes in on a final conclusion, which acts as a point-like attractor. A class two computation repeats itself, cycling around on an attractor that’s a smooth, closed hypersurface in state space. Class three processes are very nearly random and have fuzzy attractors filling their state space. Most relevant for the analysis of mind are our elaborate class four trains of thought, flickering across their attractors like fish around tropical reefs.
If you happen to be faced with a problem that actually has a definite solution, your thought dynamics can in fact can take on a class one form, closing in on the unique answer. But life’s more galling problems are in fact insoluble, and grappling with a problem like this is likely to produce a wretched class two cycle.
Taking a fresh look at a thought loop gets you nowhere. By way of illustration, no matter what brave new seed pattern you throw into a scroll-generating CA, the pattern will always be eaten away as the natural attractor of the loop reasserts itself.
So how do we switch away from unpleasant thoughts? The only real way to escape a thought loop is to shift your mind’s activity to a different attractor, that is, to undergo a chaotic bifurcation, as I put it in section 2.4. We change attractors by altering some parameter or rule of the system so as to move to a different set of characteristic behaviors. With regard to my thoughts, I see two basic approaches: reprogramming or distraction.
Reprogramming is demanding and it takes time. Here I try to change the essential operating rules of the processes that make my negative thought-loops painful. Techniques along these lines include: learning to accept irremediable situations just as they are; anticipating and forestalling my standard emotional reactions to well-known recurring trigger events; letting go of my expectations about how the people around me ought to behave; and releasing my attachment to certain hoped-for outcomes. None of these measures comes easily.
I always imagine that over time, by virtue of right thinking and proper living, I’ll be able to inculcate some lasting changes into my synapses or neurotransmitters. I dream that I’ll be able to avoid the more lacerating thought-loops for good. But my hard-won equilibrium never lasts. Eventually I fall off the surfboard into the whirlpools.
Our society periodically embraces the belief that one might attain permanent happiness by taking the right drugs—think of psychedelics in the sixties and antidepressants in the Y2K era. Drugs affect brain’s computations, not by altering the software, but by tweaking the operational rules of the underlying neural hardware. The catch is that, given the unpredictable nature of class four computation, the effects of drugs can be different from what they’re advertised to be. At this point in my life, my preference is to get by without drugs. My feeling is that the millennia-long evolution of the human brain has provided for a rich enough system to produce unaided any state I’m interested in. At least in my case, drugs can in fact lead to a diminished range of possibilities. This said, I do recognize that, in certain black moods, any theorizing about attractors and computation classes becomes utterly beside the point, and I certainly wouldn’t gainsay the use of medication for those experiencing major depression.
Distraction is an easier approach. Here you escape a problem by simply forgetting about it. And why not? Why must every problem be solved? Your mind’s a big place, so why limit your focus to its least pleasant corners? When I want to come to my senses and get my attention away from, say, a mental tape of a quarrel, I might do it by hitching a ride on a passing thought train. Or by paying attention to someone other than myself. Altruism has its rewards. Exercise, entertainment, or excursions can help change my mood.
Both reprogramming and distraction reduce to changing the active attractor. Reprogramming myself alters the connectivity or chemistry of my brain enough so that I’m able to transmute my thought loops into different forms with altered attractors. Distracting myself and refocusing my attention shifts my conscious flow of thoughts to a wholly different attractor.
If you look at an oak tree and a eucalyptus tree rocking in the wind, you’ll notice that each tree’s motion has its own distinct set of attractors. In the same way, different people have their own emotional weather, their particular style of response, thought, and planning. This is what we might call their sensibility, personality, or disposition.
Having raised three children, it’s my impression that, to a large degree, their dispositions were fairly well fixed from the start. For that matter, my own basic response patterns haven’t changed all that much since I was a boy. One’s mental climate is an ingrained part of the body’s biochemistry, and the range of attractors available to an individual brain is not so broad as one might hope.
In one sense, this is a relief. You are who you are, and there’s no point agonizing about it. My father took to this insight in his later life and enjoyed quoting Popeye’s saying: “I yam what I yam.” Speaking of my father, as the years go by, I often notice aspects of my behavior that remind me of him or of my mother—and I see the same patterns yet again in my children. Much of one’s sensibility consists of innate hereditary patterns of the brain’s chemistry and connectivity.
In another sense, it’s disappointing not to be able to change one’s sensibility. We can tweak our moods somewhat via reprogramming, distraction, or psychopharmacology. But making a radical change is quite hard.
As I type this on my laptop, I’m sitting at my usual table in the Los Gatos Coffee Roasting cafe. On the cafe speakers I hear Jackson Browne singing his classic road song, “Take It Easy,” telling me not to let the sound of my own wheels drive me crazy. It seems to fit the topic of thought loops, so I write it in. This gift from the outside reminds me that perhaps there’s a muse bigger than anything in my own head. Call it God, call it the universe, call it the cosmic computation that runs forward and backward through all of spacetime.
Just now it seems as if everything’s all right. And so what if my exhilaration is only temporary? Life is, after all, one temporary solution after another. Homeostasis.
• See book's home page for info and buy links.
• Go to web version's table of contents.
4.4: “I Am”
In section 4.1: Sensational Emotions, I observed that our brain functions are intimately related to the fact that we have bodies living in a real world, in section 4.2: The Network Within, I discussed how responses can be learned in the form of weighted networks of neural synapses, and in section 4.3: Thoughts as Gliders and Scrolls, I pointed out that the brain’s overall patterns of activation are similar to the gliders and scrolls of CAs.
In this section I’ll talk about who or what is experiencing the thoughts. But before I get to that, I want to say a bit about the preliminary question of how it is that a person sees the world as made of distinct objects, one of which happens to be the person in question.
Thinking in terms of objects gives you an invaluable tool for compressing your images of the world, allowing you to chunk particular sets of sensations together. The ability to perceive objects isn’t something a person learns; it’s a basic skill that’s hardwired into the neuronal circuitry of the human brain.
We take seeing objects for granted, but it’s by no means a trivial task. Indeed, one of the outstanding open problems in robotics is to design a computer vision program that can take camera input and reliably pick out the objects in an arbitrary scene. By way of illustration, when speaking of chess-playing robots, programmers sometimes say that playing chess is the easy part and recognizing the pieces is the hard part. This is initially surprising, as playing chess is something that people have to laboriously learn—whereas the ability to perceive objects is something that people get for free, thanks to being born with a human brain.
In fact, the hardest tasks facing AI involve trying to emulate the fruits of evolution’s massive computations. Putting it a bit differently, the built-in features of the brain are things that the human genome has phylogenetically learned via evolution. In effect evolution has been running a search algorithm across millions of parallel computing units for millions of years, dwarfing the learning done by an individual brain during the pitifully short span of a single human life. The millennia of phylogenetic learning are needed because it’s very hard to find a developmental sequence of morphogens capable of growing a developing brain into a useful form. Biological evolution solved the problem, yes, but can our laboratories?
In limited domains, man-made programs do of course have some small success in recognizing objects. Machine vision uses tricks such as looking for the contours of objects’ edges, picking out coherent patches of color, computing the geometric ratios between a contour’s corners, and matching these features against those found in a stored set of reference images. Even so, it’s not clear if the machine approaches we’ve attempted are in fact the right ones. The fact that our hardware is essentially serial tends to discourage us from thinking deeply enough about the truly parallel algorithms used by living organisms. And using search methods to design the parallel algorithms takes prohibitively long.
In any case, thanks to evolution, we humans see the world as made up of objects. And of all the objects in your world, there’s one that’s most important to you: your own good self. At the most obvious level, you pay special attention to your own body because that’s what’s keeping you alive. But now let’s focus on something deeper: the fact that your self always seems surrounded by a kind of glow. Your consciousness. What is it?
At first we might suppose that consciousness is a distinct extra element, with a person then being made of three parts:
• The hardware, that is, the physical body and brain.
• The software, including memories, skills, opinions, and behavior patterns.
• The glow of consciousness.
What is that “glow” exactly? Partly it’s a sense of being yourself, but it’s more than that: It’s a persistent visceral sensation of a certain specific kind; a warmth, a presence, a wordless voice forever present in your mind. I think I’m safe in assuming you know exactly what I mean.
I used to be of the opinion that this core consciousness is simply the bare feeling of existence, expressed by the primal utterance, “I am.” I liked the fact that everyone expresses their core consciousness in the same words: “I am. I am me. I exist.” This struck me as an instance of what my ancestor Hegel called the divine nature of language. And it wasn’t lost on me that the Bible reports that after Moses asked God His name, “God said to Moses, ‘I AM WHO I AM’; and He said, “Thus you shall say to the sons of Israel, I AM has sent me to you.’ ” I once discussed this with Kurt Gödel, by the way, and he said the “I AM” in Exodus was a mistranslation! Be that as it may. I used to imagine my glow of consciousness to be a divine emanation from the cosmic One, without worrying too much more about the details.84
But let’s give universal automatism a chance. Might one’s glow of consciousness have some specific brain-based cause that we might in turn view as a computation?
In the late 1990s, neurologist Antonio Damasio began making a case that core consciousness results from specific phenomena taking place in the neurons of the brain. For Damasio, consciousness emerges from specific localized brain activities and is indeed a kind of computation. As evidence, Damasio points out that if a person suffers damage to a certain region of the brain stem, their consciousness will in fact get turned off, leaving them as a functional zombie, capable of moving about and doing things, but lacking that glowing sense of self.
Figure 97: The Self, the World, and the Self Experiencing at the World
Your brain produces a “body-self” image on the left, an image of the world on the right, and, in the center, an image of the self interacting with the world. Damasio feels that consciousness lies in contemplating the interaction between the self and the world.
Damasio uses a special nomenclature to present his ideas. He introduces the term movie-in-the-brain to describe a brain’s activity of creating images of the world and of the body. The image of the body and its sensations is something that he calls the proto-self And he prefers the term core consciousness to what I’ve been calling the “I am” feeling.
Damasio believes that core consciousness consists of enhancing the movie-in-the-brain with a representation of how objects and sensations affect the proto-self. Putting it a little differently, he feels that, at any time, core consciousness amounts to having a mental image of yourself interacting with some particular image. It’s not enough to just have a first-order image of yourself in the world as an object among other objects—to get to core consciousness, you go a step beyond that and add on a second-order representation of your reactions and feelings about the objects you encounter. Figure 97 shows a shorthand image of what Damasio seems to have in mind.
In addition to giving you an enhanced sense of self, Damasio views core consciousness as determining which objects you’re currently paying attention to. In other words, the higher-order process involved in core consciousness has two effects: It gives you the feeling of being a knowing, conscious being, and it produces a particular mental focus on one image after another. The focusing aspect suggests that consciousness is being created over and over by a wordless narrative that you construct as you go along. Damasio describes his theory as follows:
As the brain forms images of an object—such as a face, a melody, a toothache, the memory of an event—and as the images of the object affect the state of the organism, yet another level of brain structure creates a swift nonverbal account of the events that are taking place in the varied brain regions activated as a consequence of the object-organism interaction. The mapping of the object-related consequences occurs in the first-order neural maps representing proto-self and object; the account of the causal relationship between object and organism can only be captured in second-order neural maps....One might say that the swift, second-order nonverbal account narrates a story: that of the organism caught in the act of representing is own changing state as it goes about representing something else. But the astonishing fact is that the knowable entity of the catcher has just been created in the narrative of the catching process....
Most importantly, the images that constitute this narrative are incorporated in the stream of thoughts. The images in the consciousness narrative flow like shadows along with the images of the object for which they are providing an unwitting, unsolicited comment. To come back to the metaphor of the movie-in-the-brain, they are within the movie. There is no external spectator....
The process which generates...the imaged nonverbal account of the relationship between object and organism has two clear consequences. One consequence ... is the subtle image of knowing, the feeling essence of our sense of self; the other is the enhancement of the image of the causative object, which dominates core consciousness. Attention is driven to focus on an object and the result is saliency of the images of that object in mind.85
Rephrasing the quote from Damasio, it seems that he views core consciousness as arising in the context of the following sequence:
• Immersion. You are active in the world.
• Seeing objects. You distinguish separate objects in the world, including your body.
• Movie-in-the-brain You have an ongoing mental model of the world. The movie-in-the-brain includes images of the world’s objects and an image of your body.
• Proto-self. Your image of your body differs from an image of an object in that your image of your body includes images of your sensations and current mental contents. This rich image is the proto-self.
• Feelings. You automatically and continually enhance the movie-in-the-brain by adding in representations of the protoself’s interactions with objects. These second-order representations are what we call feelings.
• Core consciousness. The act of continually forming feelings is part of what we mean by consciousness. At any given time, core consciousness is based on your feelings about a small group of images. Core consciousness highlights those particular images, which accounts for your current focus of attention.
• Empathy. You enhance your images of other people with representations of their feelings.
Empathy, in other words, is your awareness that your fellows are conscious, too. It’s possible, and not even unusual, to have consciousness without empathy—consider, for example, the all-too-common moments when one regards one’s rivals or even whole categories of people as soulless automata, as unreasoning animals, as bacilli that walk on two legs.
The thing that I find particularly striking about Damasio’s explanation of consciousness is that, being a neurologist, he makes serious efforts to identify it with a type of brain activity. This said, brain science is very much in its infancy, and no firm conclusions have been reached. But just to give the flavor of what we can expect fairly soon, Figure 98 indicates what it might be like to have a physiological brain model of the movie-in-the-brain, the protoself, feelings, and consciousness.
Figure 98: Toward a Neurology of Consciousness
(1) The thalamus gathers sensory input for images of objects. (2) The reticular formation in the brain stem gathers body sensations for the proto-self. (3) The neocortex forms the movie-in-the-brain, with images of objects and of the proto-self. (4) The cingulate region of the cortex monitors the proto-self’s reactions to the movie-in-the-brain, thereby creating core consciousness.
If indeed there’s nothing magical about consciousness, then it might as well be a type of computation. To test this notion out, I’m now going to recast Damasio’s theory in the context of the creatures in a computer game. Putting it more colorfully, I’d like to ask: What is the phenomenology of Pac-Man? Actually, since Pac-Man is externally controlled by the game’s human player, it’s really not Paccy’s phenomenology that I’m interested in.86 The real meat lies in understanding the worldview of a nonplayer character (what gamers call an NPC). What is it like to be one of the enemy ghosts who chase Pac-Man around? What is the phenomenology of a Quake monster?
You might wonder why, in the midst of an erudite philosophical discussion,
I suddenly want to start talking about something so street-level and seemingly nonintellectual as computer games. Noisy kids in strange clothes! Sex and gore! A hideous waste of time!
Academia hasn’t quite caught on to the fact that computer games represent the convergence and the flowering of the most ambitious frontier efforts of the old twentieth-century computer science: artificial intelligence, virtual reality, and artificial life.
I think I can argue that, if we create a fairly rich computer game, there’s a sense in which the game’s program-controlled creatures might be said to have core consciousness and empathy. I’ll illustrate my argument in Figure 99.
Left to right and top to bottom, the six images in Figure 99 represent six levels of sophistication.
movie-in-the-brain with proto-self
I’ll say a bit about the six levels.
Immersion. In a computer game, we model a virtual world complete with an artificial physics, and objects such as walls, food pellets, and treasures. And we place our artificial life-forms, that is, our game creatures, inside this world.
Seeing objects. In most games, programmers dodge the tricky issue of having the game creatures be able to pick out the objects in the virtual world. Rather than laboriously endowing our creatures with simulated vision and an object-recognition program, we flatly tell the creatures about the objects in their world by giving them access to a master list of the toy world’s objects, nicely identified. Movie-in-the-brain. Each of the computer-controlled game creatures has an individual update method. This method is coded in such a way that the creature can take into account the master list of all the world objects and choose a course of action accordingly.
Proto-self. It’s necessary that a creature can distinguish itself from others. If the creature is to dodge bullets, for instance, it needs to be able to measure bullets’ distances from its own body. If the creature is to bounce off other creatures, it needs to be able to determine who’s who. This is a simple matter in object-oriented programming languages like C++ and Java, both of which have a convention that when you write code to describe a given creature’s behavior, the word this is understood to refer to
Figure 99: Consciousness and Empathy for Computer Game Creatures
the creature itself. Another relevant feature of object-oriented computer languages is that certain internal variables can be designated as private. A creature will have full access to its own private variable values, but it may well lack any means of accessing the private variable values of other creatures. This illustrates the notion that a creature’s proto-self image is richer than its image of some other creature or object.
Feelings. We can equip a creature with an evaluation method that assigns positive or negative utility values to other creatures. A dangerous enemy, for instance, might have a value of negative three, whereas a piece of food could have a weight of positive two. A creature’s update method is enriched by having access to these numerical “feelings” about the other objects. In terms of programming this, we might suppose that if the world contains, say, five distinct kinds of creature, then a creature’s personal feelings are summarized in a five-element array that matches a numerical utility to each of the five creature types. Another aspect of the creature’s feelings can be one that tracks how well it’s doing. It’s common for game creatures to continually update internal score and health variables.
Core consciousness. We can imagine equipping a creature with some adaptive artificial intelligence by which it adjusts its behavior according to the situation. For instance, a creature might change the numerical values in its array of feelings about other creatures. Perhaps when it has low health it becomes more fearful and assigns its enemies a more negative utility value. When it has a low score, it might become more acquisitive and raise the utility value of food. If its health and score are both high, it might become more aggressive about attacking its enemies. So now the creature is “dealing with feelings”. The focusing aspect of core consciousness can be modeled by having the creature give greater weight to the utility value of the closest other creature, effectively paying more attention to it.
Empathy. To be really skillful, a creature might let its update method guess at its opponents’ motions in advance. At the most obvious level, the creature could look at an opponent’s current position and velocity, and estimate where it will be in the near future. But at a higher level, a creature might build up an educated guess about the feeling—like utility weights being used by its opponents. And then the creature would be that much better at simulating in advance the upcoming actions of its opponents. At a still higher level, a creature could outfox its opponents by figuring out the internal mechanisms of their update methods. In this situation, the creature is enhancing its images of other creatures with images of their feelings and with images of their core consciousness. Thus empathy.
So it seems that, if we adopt Damasio’s line of thought, a well-programmed game creature is not only conscious, it has empathy! A dippy little computer game creature is like a person. Why does this conclusion feel so utterly wrong?
A first issue is that a programmed agent on the screen doesn’t have a physical body. But suppose we were able to put these kinds of artificial minds inside of robotic bodies? What then?
A second issue is that you might persist in claiming that a program or a robot could never have the special inner glow that you sense as part of your consciousness. But maybe the glow is an accidental and inessential phenomenon. Maybe a sufficiently advanced program or robot could behave in every way like a conscious being, and even have thought processes in every way similar to our own, but even so it might not share the same visceral glow.
In discussing this type of objection to computer models of the mind, philosophers often speak of qualia, which are meant to be the ineffable, subjective sensations we experience. But if it makes no observable difference whether a robot has or doesn’t have the glow qualia, than what difference do the qualia make? And, come to think of it, how can you be sure that a robot might not experience glow qualia, perhaps in the form of some localized electrical field?
There is a tendency—which I fully understand—to think that no possible robot could ever be conscious, simply because it’s not human. But maybe this is an unreasonable prejudice. It could relate to the fact that an explained magic trick seems like no trick at all. We want to believe ourselves to be magical, spiritual beings, and it’s disillusioning to suppose that we might after all be reducible to some massive web of neural connections.
A final issue with my argument for the consciousness of computer game characters is that the actually existing virtual creatures are too simple to be regarded as conscious in any but the most trivial and limited sense of the word. After all, I normally take being conscious to mean having a fairly rich kind of internal mental life.
This objection has real weight. Perhaps the weakest link in my modeling of consciousness lies in having a computer critter’s update method represent the movie-in-the-brain stage. After all, a creature might have an update method that does no work at all. Is it really fair to compare this to a human brain’s intricate modeling of the world? Computer game creatures get the movie-in-the-brain for free because they live within the matrix of a program simulating its world. The situation is analogous to the fact that computer game creatures don’t grow their own bodies, as living organisms do. They lack morphogenesis in two senses: They don’t generate their bodies, and they don’t generate their own images of the world.
If the truth be told, I’m willing to admit that when I say a Pac-Man ghost is conscious, I’m just playing a language game. But my game has a point, this being to illustrate that we have no a priori reason for denying that a computer program could exhibit a mental life akin to a human’s—even though, in reality, we’re nowhere near that level yet.
My feeling is that, on the one hand, it is possible in principle to build or evolve humanlike robots, but that, on the other hand, in practice we won’t manage to do so anytime soon. More precisely, on the basis of a numerical argument, which I’ll explain in section 4.6: The Mind Recipe, I don’t see human technology as creating truly humanlike programs or robots any time much before the year 2100—and with no ironclad guarantee of success after then.
In a way, my talk about intelligent robots is a red herring. For me, the deeper issue is to understand, here and now, the workings of my own mind. The real point of this chapter is that our thoughts can usefully be thought of as computations, even though we can’t in fact produce detailed descriptions or emulations of what a brain does. Why is it useful to view our thoughts as computations? Because this gives us access to a new terminology in which to discuss our mental lives.
Figure 100: Is This Dog Conscious?
Pitch is famous! And there’s a picture of the spot on her back in section 3.2: The Morphogenesis of a Brindle Cow. I like how a dog’s facial expression boils down to three black dots.
The value of clear terminology cannot be underestimated. “Distinguo,” as Aristotle liked to say, “I make distinctions.” To a large extent, philosophical investigations consist of figuring out what you’re talking about. Thanks to the analysis I’m presenting here, I can make the rather precise claims that our brain activities are deterministic, class four, unpredictable, capable of the process that we’re equating with consciousness, and (at least for now) unfeasible to emulate on PCs.
What about the brains of lesser organisms? Do they support consciousness? A snail probably isn’t at the seeing objects stage. For a snail, the world is a continuum of sensory inputs, reacted to in real time. A dog can perceive objects, but I’m not sure if a dog has movie-in-the-brain or not, maybe only fleetingly (Figure 100). In The Feeling of What Happens, Damasio talks about some brain-damaged people who have movie-in-the brain but not proto-self, so a number of intermediate levels seem possible.
At this point I want to recast Damasio’s hierarchy in a fresh way so as to investigate a familiar endless regress. I’ll label Damasio’s levels with the numbers zero, one, and two—and then I’ll ask about the number levels after that. We’ll see that there’s a certain sense in which consciousness is infinite.
In particular, I’ll regard having thoughts as level zero, and becoming aware of the thoughts as level one. In Damasio’s terms, the movie-in-the-brain is level zero, and having feelings about the movie is level one. Damasio’s notion of core consciousness involves reaching level two, where one thinks about one’s thoughts.
Now note that once you reach level two, you’re free to think about consciousness itself, that is, to think about thinking about thinking. This moves you to level three. And then you can think about thinking about consciousness for level four, and so on, out through a potentially infinite series of levels, as illustrated in Figure 101.
I would say the first frame of my Wheelie Willie cartoon corresponds to the level zero movie-in-the-brain. And the second frame represents the advent of level one feelings. The endless series is set into motion only in the third frame, which represents level two consciousness and the act of thinking about thinking. Once you have this second-order thinking, you get third, fourth, fifth, and all the higher orders—at least up to the point where you get confused or bored.
We experience the higher levels, when unpleasant, as self-consciousness and irony, or, when pleasant, as maturity and self-knowledge.
In other words, the advent of consciousness introduces a dynamic that leads to a potential infinity. This is reasonable and pleasing, given that it feels natural to think of the mind as being infinite.
If you stand back, as we’re now doing, and schematically imagine running the series right out through all the natural numbers, you get a kind of enlightenment that is, however, only relative, as right away you can get to a level “infinity plus one.” The real enlightenment would be more like reaching the inconceivable Absolute Infinite that the theologically inclined mathematician Georg Cantor spoke of.
Figure 101: Wheelie Willie Thinks of Infinity and of Nothing
In 1978 I drew this cartoon of an infinite regress leading to an empty mind. I drew it the day I came home from getting fired from my first job as a math professor. Drawing it cheered me up. At first Wheelie Willie is bummed, with a patched tire. And then he gets into his head and sees infinity, which makes him happy and leaves him with a clear, empty mind. At the end the ants are crawling on him. He’s no longer a respected professor, just an ordinary hippie without a job. Maybe he’ll write a science-fiction novel!
Set theory is the branch of mathematics that concerns itself with levels of infinity, and this is the field in which I wrote my Ph.D. thesis at Rutgers University in that most sixties of years, 1972. My experience was that, when I thought hard enough about absolute infinity, I’d end up thinking about nothing at all, bringing me back to very start of the process, to immersion in the world with an empty mind with no subject-object distinctions, simply experiencing the great reality movie in and of itself, with me an integral part, filled with a visceral pleasure by the simple awareness of consciousness.
I can easily imagine a reader wondering what that last sentence is supposed to mean and what, in particular, it has to do with the notion of viewing core consciousness as a type of computation. That’s what I was supposed be talking about, right? Computation, not enlightenment.
Well, I have this abiding fondness for mysticism, coupled with a lazy man’s hope of finding a royal-road shortcut to wisdom. So it’s easy for me to flip over to guru-speak. But there’s no real conflict between mysticism and universal automatism. The individual self is simply one aspect of the computations that fill the world.
I am doing it
the it I am doing is
the I that is doing it
the I that is doing it is
the it I am doing
it is doing the I that am doing it
I am being done by the it I am doing
it is doing it
—R. D. Laing87
In section 4.6: The Mind Recipe, I’ll look at how and when a conscious machine might be evolved, and in section 4.8: Quantum Soul, I’ll consider the possibility that consciousness is, after all, quite distinct from ordinary computation. But first I want to talk about a less ambitious modeling of a human mind: the lifebox.
• See book's home page for info and buy links.
• Go to web version's table of contents.
4.5: The Lifebox
One of the most venerable dreams of science fiction is that people might become immortal by uploading their personalities into some kind of lasting storage. Once your personality is out of your body in a portable format, it could perhaps be copied onto a fresh tank-grown blank human body, onto a humanoid robot, or, what the heck, onto a pelican with an amplified brain. Preserve your software, the rest is meat!
In practice, copying a brain would be very hard, for the brain isn’t in digital form. The brain’s information is stored in the geometry of its axons, dendrites, and synapses, in the ongoing biochemical balances of its chemicals, and in the fleeting flow of its electrical currents. In my early cyberpunk novel Software, I wrote about some robots who specialized in extracting people’s personality software—by eating their brains. When one of my characters hears about the repellent process, “[his] tongue twitched, trying to flick away the imagined taste of the brain tissue, tingly with firing neurons, tart with transmitter chemicals.”88
In this section I’m going to talk about a much weaker form of copying a personality. Rather than trying to exactly replicate a brain’s architecture, it might be interesting enough to simply copy all of a person’s memories, preserving the interconnections among them.
We can view a person’s memory as a hyperlinked database of sensations and facts. The memory is structured something like a Web site, with words, sounds, and images combined into a superblog with trillions of links.
I don’t think it will be too many more years until we see a consumer product that makes it easy for a person to make a usable copy of their memory. This product is what I call a lifebox.89
My idea is that your lifebox will prompt you to tell it stories, and it will have enough low-level language recognition software to be able to organize your anecdotes and to ask you follow-up questions. As you continue working with your lifebox, it builds up a database of the facts you know and the tales you spin, along with links among them. Some of the links are explicitly made by you, others will be inferred by the lifebox software on the basis of your flow of conversation, and still other links are automatically generated by looking for matching words.
And then what?
Your lifebox will have a kind of browser software with a search engine capable of returning reasonable links into your database when prompted by spoken or written questions from other users. These might be friends, lovers, or business partners checking you out, or perhaps grandchildren wanting to know what you were like.
Your lifebox will give other people a reasonably good impression of having a conversation with you. Their questions are combed for trigger words to access the lifebox information. A lifebox doesn’t pretend to be an intelligent program; we don’t expect it to reason about problems proposed to it. A lifebox is really just some compact digital memory with a little extra software. Creating these devices really shouldn’t be too hard and is already, I’d say, within the realm of possibility—it’s already common for pocket-size devices to carry gigabytes of memory, and the terabytes won’t be long in coming.
I discussed the lifebox at some length in my Y2K work of futurology, Saucer Wisdom, a book that is framed in terms of a character named Frank Shook who has a series of glimpses into the future—thanks to some friendly time-traveling aliens who take him on a tour in their tiny flying saucer. (And, no, I’m not a UFO true believer, I just happen to think saucers are cute and enjoyably archetypal.) Here’s some quotes from the book.
The lifebox is a little black plastic thing the size of a pack of cigarettes and it comes with a lightweight headset with a pinhead microphone, like the kind that office workers use. You can use your lifebox to create your life story, to make something to leave for your children and grandchildren.
Frank watches an old man using a lifebox. His name is Ned. White-haired Ned is pacing in his small backyard—a concrete slab with some beds of roses—he’s talking and gesturing, wearing the headset and with the lifebox in his shirt pocket. The lifebox speaks to him in a woman’s pleasant voice.
The marketing idea behind the lifebox is that old duffers always want to write down their life story, and with a lifebox they don’t have to write, they can get by with just talking. The lifebox software is smart enough to organize the material into a shapely whole. Like an automatic ghostwriter.
The hard thing about creating your life story is that your recollections aren’t linear; they’re a tangled banyan tree of branches that split and merge. The lifebox uses hypertext links to hook together everything you tell it. Then your eventual audience can interact with your stories, interrupting and asking questions. The lifebox is almost like a simulation of you.
To continue his observations, Frank and his friends skip forward in time until past when Ned has died and watch two of Ned’s grandchildren play with one of the lifebox copies he left, as shown in Figure 102.
Frank watches Ned’s grandchildren: little Billy and big Sis. The kids call the lifebox “Grandpa,” but they’re mocking it, too. They’re not putting on the polite faces that kids usually show to grown-ups. Billy asks the Grandpa-lifebox about his first car, and the lifebox starts talking about an electric-powered Honda and then it mentions something about using the car for dates. Sis—little Billy calls her “pig Sis” instead of “big Sis”—asks the lifebox about the first girl Grandpa dated, and Grandpa goes off on that for a while, and then Sis looks around to make sure Mom’s not in earshot. The coast is clear so she asks some naughty questions. “Did you and your dates do it? In the car? Did you use a rubber?” Shrieks of laughter. “You’re a little too young to hear about that,” says the Grandpa-lifebox calmly. “Let me tell you some more about the car.”
Frank then skips a little further into the future, and he finds that lifeboxes have become a huge industry.
People of all ages are using lifeboxes as a way of introducing themselves to each other. Sort of like home pages. They call the lifebox database a context, as in, “I’ll send you a link to my context.” Not that most people really want to spend the time it takes to explicitly access very much of another person’s full context. But having the context handy makes conversation much easier. In particular, it’s now finally possible for software agents to understand the content of human speech—provided that the software has access to the speakers’ contexts.90
Figure 102: Grandchildren with a Lifebox
Frank Shook is inside the little UFO, which is invisible to the children.
Coming back to the idea of saving off your entire personality that I was discussing at the beginning of the section, there is a sense in which saving only your memories is perhaps enough, as long as enough links among your memories are included. The links are important because they constitute your sensibility, that is, your characteristic way of jumping from one thought to the next.
On their own, your memories and links aren’t enough to generate an emulation of you. But when another person studies your memories and links, that other person can get into your customary frame of mind, at least for a short period of time. The reason another person can plausibly expect to emulate you is that, first of all, people are universal computers and, second of all, people are exquisitely tuned to absorbing inputs in the form of anecdotes and memories. Your memories and links can act as a special kind of software that needs to be run on a very specialized kind of hardware: another human being. Putting it a bit differently, your memories and links are an emulation code.
Certainly exchanging memories and links is more pleasant than having one’s brain microtomed and chemically analyzed!
I sometimes study authors’ writings or artists’ works so intensely that I begin to imagine that I can think like them. I even have a special word I made up for this kind of emulation; I call it twinking. To twink someone is to simulate them internally. Putting it in an older style of language, to twink someone is to let their spirit briefly inhabit you. A twinker is, if you will, like a spiritualistic medium channeling a personality.
Over the years I’ve twinked my favorite writers, scientists, musicians, and artists, including Robert Sheckley, Jack Kerouac, William Burroughs, Thomas Pynchon, Frank Zappa, Kurt Gödel, Georg Cantor, Jorge Luis Borges, Edgar Allan Poe, Joey Ramone, Phil Dick, and Peter Bruegel. The immortality of the great ones results from faithful twinking by their aficionados.
Even without the lifebox, if some people don’t happen to be authors, they can make themselves twinkable simply by appearing in films. Thomas Pynchon captures this idea in a passage imagining the state of mind of the 1930s bankrobber John Dillinger right before he was gunned down by federal agents outside the Biograph movie theater in Chicago, having just seen Manhattan Melodrama starring Clark Gable.
John Dillinger, at the end, found a few seconds’ strange mercy in the movie images that hadn’t quite yet faded from his eyeballs—Clark Gable going off unregenerate to fry in the chair, voices gentle out of the deathrow steel so long, Blackie...there was still for the doomed man some shift of personality in effect—the way you’ve felt for a little while afterward in the real muscles of your face and voice, that you were Gable, the ironic eyebrows, the proud, shining, snakelike head—to help Dillinger through the bushwhacking, and a little easier into death.91
The effect of the lifebox would be to make such immortality accessible to a very wide range of people. Most of us aren’t going to appear in any movies, and even writing a book is quite hard. Again, a key difficulty in writing any kind of book is that you somehow have to flatten the great branching fractal of your thoughts into a long line of words. Writing means converting a hypertext structure into a sequential row—it can be hard even to know where to begin.
As I’ve been saying, my expectation is that in not too many years, great numbers of people will be able to preserve their software by means of the lifebox. In a rudimentary kind of way, the lifebox concept is already being implemented as blogs. People post journal notes and snapshots of themselves, and if you follow a blog closely enough you can indeed get a feeling of identification with the blogger. And blogs already come with search engines that automatically provide some links. Recently the cell phone company Nokia started marketing a system called Lifeblog, whereby a person can link and record daily activities by using a camera-equipped cell phone.
Like any other form of creative endeavor, filling up one’s lifebox will involve dedication and a fair amount of time, and not everyone will feel like doing it. And some people are too tongue-tied or inhibited to tell stories about themselves. Certainly a lifebox can include some therapist-like routines for encouraging its more recalcitrant users to talk. But lifeboxes won’t work for everyone.
What about some science-fictional instant personality scanner, a superscanner that you wave across your skull and thereby get a copy of your whole personality with no effort at all? Or, lacking that, how about a slicer-dicer that purees your brain right after you die and extracts your personality like the brain-eaters of Software? I’m not at all sure that this kind of technology will ever exist. It’s hard to infer the high levels from the low. And the brain’s low levels may prove too delicate to capture.
I like the idea of a lifebox, and I have vague plans to try to make one for myself. I envision a large database with all my books, all my journals, and a connective guide-memoir—with the whole thing annotated and hyperlinked.
And I might as well throw in some photographs—I’ve taken thousands over the years. And it should be feasible to endow my lifebox with interactive abilities; people could ask it questions and have it answer with appropriate links and words. My finished lifebox might take the form of a Web site, although then there’d be the thorny question of how to get any recompense for the effort involved. A commercial alternative would be to market it as a set of files on a portable data storage device of some kind. Rudy’s Lifebox—my personal pyramid of Cheops.
But I don’t really think the lifebox would be me. Without some radically more powerful software, it would just be another work of art, not so different from a bookshelf of collected works.
In the next section I’ll examine the question of whether humans will ever manage to design the kind of software that could turn a lifebox into a conscious mind.
• See book's home page for info and buy links.
• Go to web version's table of contents.
4.6: The Mind Recipe
Suppose you have a lifebox copy of what you know. Now you’d like to animate it so as to have an artificial version of yourself. How would you go about creating a humanlike intelligence?
The easy way to create human minds is, of course, to raise children. Let biology do the computing! And, if you’re fortunate enough to have children, you can try to teach them everything you know, thus achieving a kind of mind replication. But for some reason many of us find it interesting to think about emulating human beings with software running on machines like souped-up desktop computers or, geekier yet, on high-tech devices inside robot bodies. We might pause here and wonder why this seems like a reasonable idea—because maybe it isn’t.
In past eras, people have imagined creating humans from mud and magic, from clockwork machinery, from magnets and motors, or from telephone switching circuits. Why give serious credence to today’s dreams of chip-based humanoids?
Some would indeed argue that there is something uniquely nondeterministic about a human mind, something that must elude any digital emulation. The feeling here is that roboticists are living in a fantasy world, hypnotized by their toys, blind to the richness of their own mental lives. I’m not entirely unsympathetic to this position. But for now, let’s push ahead with the notion of modeling the mind with personal computer software and see where the investigation leads.
One thing electronic computers have going for them is universality. If what my brain does is to carry out computation-like deterministic processes, then in principle there ought to be a digital electronic computer program that can emulate it. It’s presently not feasible to make such machines, but perhaps someday we’ll have them. A computation is just a computation, and our PCs are doubling their power every year or so.
But remember that our brain’s design was computed over millions of years of evolution involving millions of parallel lives. As I’ll be explaining here, my guess is that it’ll take at the very least until 2100 before machines can catch up—and maybe a lot longer than that.
In trying to produce humanlike robot brains, we have two problems to deal with: hardware and software.
The hardware problem is a matter of creating a desktop computer that has about the same computation rate and memory size as a human brain.
The software problem is to program a sufficiently powerful PC so as to behave like a person.
What I’m going to do now is to get some numerical estimates of the difficulty of the two problems and make some guesses about how soon they might be solved.
To begin with, I’ll crunch some numbers to estimate how long it might take to solve the hardware problem by creating digital computers capable of computing as fast as a human brain. Then I’ll turn to the software problem. As you might expect, the software problem is by far the harder of the two. I’ll discuss a method we might use to evolve the software for the brain—I’ll call this a mind recipe. And then I’ll estimate how long it might be until we can actually execute something like the mind recipe to produce the software for an artificial brain.
Before getting into the details, I want to review the standard American nomenclature for large numbers. As a mathematician I’m always disappointed when I see authors use expressions like “million million” or “million billion” when they could just as well be using the much more attractive words trillion and quadrillion, respectively. One possible reason journalists hesitate to use the proper words is that up until the middle of the twentieth century, America and many European countries used different naming systems for large numbers.92 So there’s some lingering confusion. But the nomenclature in Table 10 has become a standard, and it should be confidently used; the table lists the number names and the corresponding prefixes as decreed by the General Conference of Weights and Measures.
Power of ten
Table 10: Names for Large Numbers
The “Etym.” column suggests the etymology or origin for the prefixes. The unfamiliar prefixes zetta and yotta are official, as in, “our universe may have a radius of a hundred yottameters, with the largest known galaxy being some fifty zettameters across.” The idea is that zetta is like the letter Z, and that the prefixes beyond it can move backward through the alphabet, with yotta thus being like a letter Y. The xenna and watta prefixes are unofficial and may not stick. If not, watta flop!
By way of further laying the groundwork for discussing the relative power of machines and the human brain, I also need to mention that when we have a system that’s repeatedly doing something, we use the hertz unit to measure how rapidly it’s cycling. A hertz (Hz for short) is simply one cycle per second. Cycle of what? This depends on the system you’re interested in.
In the case of today’s PCs, the cycles we measure are ticks of the system clock that controls the machine’s chips. Unlike biological organisms, most electronic computers need to keep their activities in rigid synchronization, and their system clocks pulse out signals at a certain rate.
Combining our standard prefixes with the “hertz” word, we see that a gigahertz computer operates at a billion cycles per second, a terahertz machine at a trillion cycles per second, and so on.
How does a PC’s clock cycle frequency relate to the amount of computation being done?
We often discuss the speed of an electronic computer in terms of the rate at which it executes machine instructions. By a machine instruction, I mean a primitive chip-level operation such as “add the contents of memory register B to memory register A,” or “copy the contents of memory location S into memory register A.” Instead of speaking of clock cycles, it makes sense to measure a computer’s speed in instructions per second, sometimes abbreviated as IPS.
Now, in comparing hertz to IPS, we need to think about how many clock cycles a machine instruction requires. A typical instruction might take two to four clock cycles, depending on what the instruction does and the kind of design the chip has. For the purposes of this discussion, I’m just going to peg the average number of cycles per machine instruction at three. This means that a machine that executes k instructions per second is roughly the same as a machine whose system clock ticks 3 • k times per second.
Although in the early Y2K era one often heard the hertz measure when shopping for a PC, this measure is going to be less common in the future. The reason is that as we turn to more highly parallel kinds of computer architectures, our machines start executing dozens, hundreds, or even thousands of instructions per clock cycle. For these architectures, the IPS measure is more meaningful than the hertz measure. A machine with several thousand processors may be executing a million IPS even if it’s only running at a thousand hertz.
A variation on IPS is to measure a computer’s speed in so-called floating-point operations per second. A floating-point operation—called a flop for short—means doing something like adding or multiplying a pair of continuous-valued variables. I’m not talking about endlessly detailed mathematical real numbers, of course, just digitally-coded real numbers that have been rounded off into a few bytes of information. A nice thing about “flop” is that the root combines readily with the standard prefixes like giga and tera. A teraflop machine would be capable of executing a trillion floating-point operations a second. A few people insist on tacking a final “s” on the flop names, but I don’t like to do that. Megaflop, gigaflop, teraflop, petaflop, exaflop, zettaflop, yottaflop, xennaflop, wattaflop!
As our machines become more advanced, the difference between an IPS and a flop becomes less significant—executing a floating point operation is a bit more work than carrying out a more primitive instruction, so it used to be that a machine’s flop rate was normally a bit lower than the IPS rate. But today’s processors tend have machine-level instructions that carry out floating-point operations in a single instruction step. So for the rest of this section, I’ll treat IPS and flop as being more or less the same, with each of them being roughly equivalent to three hertz.
I myself have been partial to the flop nomenclature for years. In my 1980s science-fiction novels Software and Wetware, I describe some humanlike robots called boppers. Having already done some estimates of the human brain rate at that time, I had my boppers running at petaflop speed, with a new generation of exaflop models coming in.
One final measure of computer speed that one hears is MIPS, which is a million IPS, that is, a million instructions per second. To my mathematics-professor way of thinking, MIPS is just a cop-out way of avoiding using the fancy names for really big numbers. But you see it a lot.
So now let’s discuss what kind of hardware we’d need to match a human brain. Some recent estimates of the human brain’s power are summarized in Table 11.
A hundred trillion IPS
A hundred million MIPS
A hundred teraflop
Twenty billion MIPS
Three hundred billion MIPS
Three hundred petaflop, or nearly an exaflop
Table 11: Estimates of the Brain’s Rate of Computation
An IPS is an instruction per second, a MIPS is one million IPS, a flop is roughly the same as an IPS, and an IPS translates into about three hertz.
Hans Moravec is a lively roboticist who makes his home at Carnegie-Mellon University. He and his students have done a lot of work on self-driving cars. He’s the author of several brilliant books on robotics, including Robot: Mere Machine to Transcendent Mind, from which his estimate of the human brain’s power is taken. Speaking of Moravec, a mutual friend of ours once said, “Ah, Hans. He’s hard-core. I think if anyone could ever convince him that robots could never be conscious, he’d lose his will to live.” Ray Kurzweil is a Silicon Valley inventor-entrepreneur who’s developed devices for optical character recognition, speech recognition, music synthesis and more. I take his estimate of the human brain’s power from his best-selling The Age of Spiritual Machines.93
There’s a huge amount of guesswork (and perhaps outright fantasizing) in any estimate of the human brain’s computation rate. To give the flavor of how it’s done, Table 12 shows how I came up with my current three hundred petaflop or exahertz estimation for a brain-equivalent PC. And do keep in mind that every number in the table is debatable.
How soon might we expect our PCs to reach the petaflop or exahertz zone? In 1964, the engineer Gordon Moore noticed that the number of transistors per computer chip seemed to be doubling every year or two. The law extends to the speed of our computers as well, with the power of a typical desktop PC doubling every eighteen months. A little math shows that over fifteen years of Moore’s Law growth, PCs increase their power by a thousandfold. So far it’s worked. Machines of the early 1970s ran in the kilohertz range, machines of the mid-1980s ran in the megahertz range, and now in the early 2000s our machines are running in the gigahertz range. A millionfold speedup in the course of thirty years (Figure 103).
Estimate in words
Neurons per brain
A hundred billion.
Computational elements per neuron.
A neuron has an average of a thousand synapses, including the input dendrites and the branching output axon. If we view the neuron’s computing elements as consisting of its synapses along with the central neuron body itself, we still get about a thousand.
Machine instruction equivalent of updating a synapse or a neuron body
I’m imagining that we can use a few bytes of computer state for each synapse or neuron body, and that updating one of these states will involve reading some neighboring state values, adding and multiplying a few numbers, and saving the results back into the given computational element’s memory. Let’s suppose that thirty machine instructions (or flops) will be enough to update the internal state of either a synapse or the central body of a neuron.
A neuron’s computational updates per second
Typical neurons can fire ten times a second. Let’s suppose that simulating a single firing event requires ten updates to the neuron’s body and dendrites. This gives us one hundred computational updates per second.
Instructions (or floating point operations) per second
Now we multiply the previous four numbers. Neurons per brain • computational elements per neuron • instructions per element update • neuron updates per second = 30 • 10(11+3+2) = 30 • 1016 = 3 • 1017 or three hundred quadrillion instructions per second, which we can also view as three hundred petaflops.
3 • 1017
Comparable clock rate
Say there’s three clock ticks per machine instruction, and get 3 • 3 • 1017 cycles per second. I’ll round this up to 10 • 1017 to get 1018, which is a tidy quintillion ticks per second, or one exahertz.
Table 12: The Three-Hundred-Petaflop Brain as an Exahertz PC
Were Moore’s Law to continue holding good, we could get a billionfold speedup in forty-five years and reach the petaflop-to-exaflop zone of human brain hardware around the middle of the century. So does that mean we’ll have humanlike robots in the year 2045?
Not so fast. First of all, we don’t know how long Moore’s Law will hold up. As of the writing of this book, some pessimistic chip engineers are saying that we can hope for at most another fifteen years of Moore’s Law, meaning that our PCs are going to have a hard time getting much past the teraflop range. But maybe they’re wrong. People have predicted the end of Moore’s Law before, and over and over new technologies have emerged to keep things going.
Figure 103: Moore’s Law Forever?
We plot a trend whereby our computers get a thousand times as fast every fifteen years. Tracking this triumphant upward line, we get machines with hardware speeds comparable to the exaflop human brain in 2045, and in 2105 we get wattaflop machines powerful enough to evolve the software needed to make the exaflop machines actually think like human brains. But some computer engineers think the graph is going to level out to a much lower growth rate by the year 2030. In order to make this graph easy to draw, I used a linear scale on the horizontal axis and a so-called logarithmic scale on the vertical axis. This means that moving one notch on the horizontal scale adds some fixed amount (fifteen years), while moving a notch on the vertical scale multiplies by a fixed amount (a thousand.).
A more serious reason why we shouldn’t expect humanoid robots by 2045 is that, as I’ve mentioned several times, finding the correct software to emulate the brain is a very hard problem. Keep in mind that most of your brain’s programming is something you were born with—the fruit of thousands upon thousands of years of evolution. A petaflop or exaflop machine with a blank disk drive isn’t suddenly going to wake up and be like a person when you turn it on. We need to face the problem of inventing or evolving the brain emulation software to put onto the machine.
At our present state of knowledge, it appears that actually designing humanlike software is too difficult to be solved by any method other than a massive search procedure. I’m now going to loosely describe a specific way to carry out such a search. I’ll call it the mind recipe.
My mind recipe is a deterministic procedure that could lead in a finite amount of time to a computer program that acts like a human being. There’s nothing particularly impractical about my recipe, by the way. I’ve combined familiar ideas from the field of artificial intelligence in a new synthesis.
The mind recipe is a collection of simulated evolutions, woven together in a special way. The mind recipe is meant to function as a specific description of a procedure one might set in motion upon some powerful computers, with the expectation that if the recipe “cooks” long enough, a humanlike mind will result. I’ll describe the mind recipe in chunks.
• Agent architecture
• Fitness tests
• Size limit
• Judging success
Agent architecture. The most natural idea is to use Marvin Minsky’s notion of building up our robot brains in a hierarchical fashion. Each brain would consist of systems that we call agents, with the understanding that an agent can itself be made up of agents, of neural nets, or of a combination of agents and neural nets.
At the most primitive ground-level of design, we use neural nets because we already know how to work with them. We package neural nets into agents and then let higher-level agents use these existing agents. These higher-level agents can be used as components of still higher-level agents, and so on. Any agent can serve as a black box or a subroutine to be used by another agent. In any individual agent’s design, we may also use a network of neural nets to wire its component agents together.
It’s worth noting that the agent hierarchy need not be strictly linear: an agent can feed information back into one of its subagents, producing a possibly interesting or useful feedback loop.
The value of working with an agent architecture is that this allows us to break the problem into more manageable pieces such as recognizing objects, remembering events, forming plans, understanding speech, and so on—with, of course, each of these tasks having its own subtasks to be solved by lower-level agents.94
Schedule. I see the mind recipe schedule or time line being set up as a symbiotic co-evolution of various kinds of agents. The mind recipe schedule orchestrates a number of evolutionary processes in parallel arenas. We designate one arena as the master, with the goal of evolving a mind. The goals of the subsidiary arenas are to evolve specialized agents. The schedule regularly takes the best agents from each arena and makes them available as new improved components for the agents in the other arenas.
Using co-evolution and parallel arenas allows the mind recipe to divide and conquer. Right from the start the master arena tries to evolve an entire mind, but at the same time the schedule is evolving good agents for a myriad of basic tasks. The schedule can also include an ability to change the goal of a given evolutionary arena once it’s served its purpose.
Evolution. We simulate evolution in a given arena of agents as follows. We pick some fixed size and populate the arena with this many agents. So as not to waste time reinventing the wheel, we include in each evolutionary arena some variations on AI agents that represent the current best state of the art for that arena’s task. But we’ll also include a number of randomly designed agents.
In the process of simulated evolution, we measure each agent’s fitness according to some test that relates to our eventual goal. We kill off the least fit agents and replace them with variations of the most fit agents, sometimes combining one or more of the fit agents to create a new one. And then we run the fitness tests again, kill the losers, replicate the winners, and so on. Typically we keep the population size constant—always replacing each unfit agent by exactly one new agent.
Variation. During the process of our simulated evolutions, we vary the agents so as to explore our space of possibilities. In varying an agent, we can tweak it in several ways. Most simply, we can change its neuron’s input weights and threshold levels. We can also tweak an agent by varying its neural network design, that is, by changing the number of component neurons the agent uses, or by altering the connections among inputs, neurons, and outputs. A final way to vary an agent is to change which subagents or sensors it uses as inputs.
Another source of variation is that we can crossbreed two or more agents by exchanging bits of their component neural nets.
Pseudorandomizer. A genetic algorithm involves making random choices in the tweaks and in picking which pairs of agents to crossbreed. We can make our mind recipe deterministic by getting its randomness from a pseudorandom function such as Stephen Wolfram’s cellular automaton Rule 30. (But do see footnote 73).
Fitness tests. The mind recipe uses three kinds of fitness tests.
First of all we will equip it with a number of simple tutorial programs, not unlike what a person might encounter as an online course. These automated tutorials will coach the agents through various elementary tasks, with perhaps the hardest task being to learn English (or, if you prefer, some other human language). Each of the evolutionary arenas will have its own particular series of fitness tests. Once the agents in a given arena have been brought up to a certain level, the coded-up tutor embedded in the mind recipe will propose a harder task.
Secondly, the mind recipe will include as a database a large library of books, photographs, and movies, as well as batteries of quizzes about the texts and images. Our candidate mind programs must learn to understand our books, decipher our images, and respond to our movies.
The third type of fitness test will be of a more interactive nature: once the agents reach a certain level of competence, we’ll place them into a simulated virtual world and let them directly compete with one another. This brings in the subsidiary problem of running a good virtual reality simulation, but, on the whole, that seems like a more feasible problem than evolving intelligent programs. It’s already reasonable to imagine, for instance, a big computer game in which the nonplayer characters compete without any human input at all.
Compete in what way? I see the mind recipe as having different epochs. To begin with, the virtual agents can compete for scarce resources. Next they might hunt and eat one another. Then they might group together, forming a virtual economy, buying and selling things. Finally, they can begin to exchange information with each other, offering solutions to problems, or perhaps even bartering entertainment. Once the agents reach this epoch, it makes sense to let them generate their own ranking system to be used as a fitness function.
Autonomy. In his stories about robots, Isaac Asimov had a character named Susan Calvin who served as a tutor to the robots. But I am requiring that the mind recipe be autonomous. It must work without any real-time human tutoring or interaction of any kind; once it’s set in motion, it requires no further intervention at all. The mind recipe is a start-it-and-forget-it automatic process; you turn it on, walk off, come back later (maybe much later), and you’ve got a mind waiting there.
A first reason for the autonomy requirement is a practical one: no human would have the patience and the rapidity to mentor each member of an evolving race of artificially intelligent agents.
A second reason for automating the mind recipe is that then, the faster our hardware gets, the faster we can run the mind recipe. Since no humans are involved, we’re perfectly free, as our hardware makes it possible, to run the evolution a million, billion, or trillion times faster. This is a very big win.
A third reason for making the mind recipe self-contained is that then the recipe can be fully deterministic. Once the mind recipe is fully specified, the only input is what initial seed you put into the pseudorandomizer. You’ll get the exact same final agents if you run the mind recipe two times in a row on the same seed and for the same amount of time. This fact is of some philosophical interest, as it shows that a human-equivalent mind may in fact be rather concisely describable (in terms of a mind recipe).95
Figure 104: Hans Moravec’s Plot of Brain Speed and Brain Size
The horizontal axis measures the system’s memory in megabytes, and the vertical axis shows its speed in MIPS (millions of instructions per second). The systems that can easily be made to junction as universal computers are marked with an asterisk. Note that Moravec uses logarithmic scales on both axes, with each step representing a multiplication by a thousand.
Size limit. If we put no upper bound on the size of our agents’ programs, this poses a danger of evolving toward enormous lookup tables or “crib sheets” containing the answers to all the questions asked by the fitness functions.
So as to avoid having the programs cheat by hard-coding the answers to the tests, we would want to put some kind of cap on the size of the programs. We already know how to do this by having a neural net of a rather small fixed size learn to distinguish members of a very large database. The net doesn’t have room to memorize each item, so it’s forced to find higher-level ways of distinguishing its inputs. Indeed, one might even say that intelligence involves finding compact representations of large domains.
How large should the size cap be? Hans Moravec remarks that most living or artificial thinking systems have a memory size in bytes comparable to the number of instructions per second they can carry out (see Figure 104).
Using Moravec’s rule of thumb, we should expect a three-hundred-petaflop human brain program to use about three hundred petabytes of memory. And this is about right, for if we look back at Table 12, we see that a brain has a hundred trillion synapses. If we suppose that fully describing the state of a synapse takes about a thousand bytes, then we get a hundred quadrillion bytes, or a hundred petabytes. To be generous, we can set the size cap for our agents at an exabyte apiece.
Creativity. One subsidiary goal for our agents will be the ability to mentally simulate events, to form a movie-in-the-brain. Once a program can simulate the world as it is, there’s a possibility of gaining a capacity to simulate the world as it might be. And with this comes the ability to plan and anticipate. Harking back to Damasio, also note that creating a movie-in-the-brain is a key step toward core consciousness.
A further effect of mentally simulating the world is that it presents the possibility of abstract thought. An abstract thought is, I would say, any link between several sensations or images. The more images involved, the abstracter the thought.
Once you have abstract thought, why not expect creativity? A classic objection to robot intelligence is that robots act only in accord to a program, whereas humans can do creative and unexpected things.
But a point I’ve been hammering on throughout The Lifebox, the Seashell, and the Soul is that a class four computation can be both unpredictable and deterministic. Presented with a computer program allegedly equivalent to your mind, you’d in fact have no hope of skimming through the program, understanding it, and being able to predict what it’s going to do. Heck, we can’t even predict what the CA Rule 110 will do on arbitrary inputs, and it’s a program that’s only eight bits long!
Unpredictability is easy. The real problem lies in getting a computer program to produce unexpected outputs that seem interesting.
Judging success. Certainly we’d want our program to be able to give good answers to questions about all the books and movies that we included in the mind recipe. A benefit of having used this as input means that the program will have a very strong understanding of human language. It’s likely, however, that the co-evolution in virtual reality will also have led our computer minds to develop a native language of their own.
A payoff from having used the virtual reality arena is that this will have also prepared the minds to move around our physical world. We’d want to be able to “decant” the mind into a robot body and have it move around our world in a fairly natural way, possibly needing a few weeks of acculturation.
Alan Turing spoke of gauging humanlike intelligence by an imitation game: the machines try to trick us into thinking they’re people. If they regularly succeed, we might as well say they’re intelligent.
But, and this is a delicate point, even if our mind recipe leads to fully humanoid robots, we won’t be able to prove that the machines are equivalent to human minds. Nor will we even be able to prove that these evolved minds might not at some point begin acting in a very unreasonable or inconsistent way.96 Lest this make robots seem dangerous, note that this is exactly the same situation we’re in vis-a-vis other humans. You can never be sure if another person will remain consistent and rational. Think of your relatives!
Runtime. As I mentioned before, given our present skills, humanoid software can only be designed by an evolutionary search procedure along the lines of the mind recipe. Let’s try to estimate how long might such a search take.
That is, I’d like to get a numerical estimate of how soon we might be able to evolve software equivalent to the human mind by using something like the mind recipe. But at present the mind recipe is still a bit vague, and I don’t see a simple way to make an estimate of how long it might take to bear fruit.
So I’m going to use a different approach to getting an estimate of how long it might take to evolve humanlike software: I’m going to propose directly simulating human evolution, starting with some blank, but human-size brains. That is, I’ll work with a sizable population of agents, each of whom has the three-hundred-petaflop power of a human mind, and I’ll imagine evolving these agents over a substantial period of time. To be more precise, I’ll simulate a million agents over a million years.
I’ll also need to simulate a nice virtual reality for these agents to live in, but my sense is that it’ll be enough to focus only on the computation needed to simulate all those human-size brains. As long as we stick to a fairly coarse level of simulation, running a virtual reality isn’t anywhere near as hard as simulating minds. The real issue is simulating a million years of life for a million human-size brains. Table 13 outlines my calculation that a wattaflop computer could do this in a year of continuous runtime.
How long would Moore’s Law take to get us to the wattaflop level? Supposing that we’re roughly at the gigaflop level, we’d need a speedup factor of 1021. This amounts to chaining together seven speedups of a thousandfold each. So if a Moore’s Law increase of a thousandfold takes fifteen years, we’ll need a mere ninety-five years to hit the wattaflop level. Call it a century from now, or 2100.
Actually, there’s no reason we couldn’t run the million simulated brains on a million different machines that communicate over the Net. In this case, we could reach our goal thirty years sooner. That is, assuming sixty-five years of Moore’s Law, we’d reach the yottaflop level, at which time we could run a million yottaflop machines for year—which would be as good as waiting thirty more years to run one wattaflop machine for a year.
Yottaflop, wattaflop—maybe this is getting ridiculous. How long can we really expect Moore’s Law to hold up? Earlier I mentioned an estimate that the universe has performed 10120 computations so far. So if we get really demanding and ask for a machine that can simulate the whole universe in a second, we’d need to go to 10120 computations per second from our current rate of roughly 109, which means a speedup factor of 10111. Well, if we turn the Moore’s Law crank one more time, out comes a prediction that our machines will run this fast in about five hundred years. So, come 2500, if not sooner, a desktop machine can in one second compute as much as the entire universe to date. And fifteen years after that, the computer can simulate itself running a thousand times as fast as it actually runs! Now if you believe that, I have some crystal-treated healing water to sell you, also a plot of land on the Moon, also the collected works of Plato intaglio-printed on a ZigZag cigarette paper for easy ingestion. And for an extra $100,000, I’ll make you immortal by cryogenically freezing your nose.
Estimate in words
A million population slots. This means a million at any given time, not a million in all, with dying individuals being replaced by newborns. A million is on the low side, but remember that Earth’s population used to be much smaller.
Number of years of evolution
A million years. Not a very long time on the evolutionary scale, but it might be enough for our simulated evolution. We can let our simulated creatures produce a new generation once every simulated year, a faster turnover rate than the twenty or so years per generation in the real human world.
Total number of brain years needed to simulate human evolution
Population • years of evolution =
1012 brain years
Target rate to simulate this on one machine in one year
To compute 1012 brain years in one year, the machine needs to run 1012 times as fast as a brain-simulating machine, which we know from the last table to run at 3 • 1017 instructions per second, so we’ll need 3 • 1029 instructions per second, which we might as well round up to 1030, a tidy million yottaflop, which could also be called a wattaflop.
13: A One-Year Simulation of Human Evolution on a Wattaflop Machine
My point is that at some point Moore’s Law really does have to poop out. Obviously this will happen before we see the true and complete history of the entire universe being simulated in every detail in one second on a personal computer. And it’s also unlikely that we’ll ever get yottaflop machines capable of ganging up to perform a one-year emulation of the human race’s evolution. For that matter, even the millionfold speedup to approach the power of the human brain could be problematic. As an example of the engineering problems that arise, if a computer is based on an electronic circuit that oscillates at an exahertz rate, the circuit will give off hard X-rays that disrupt the functioning of the circuit.
But on the optimistic side, even if the chip engineers hit a wall, this doesn’t necessarily mean we’ll never see an exaflop desktop device. Massive parallelism could save the day: processors could become so cheap and tiny that home computers could have thousands or even millions of them. Or perhaps the current chip architecture of etched silicon might be replaced by something more exotic like carbon nanotubes, organic molecules, or computational plastics. Or we may turn more and more to biologically designed systems. It could be that we come full circle, and the desktop computer of the year 3000 is a brain in a jar.
Still speculating on how to get really fast computers, it may also be that quantum computation will start pulling rabbits out of its hat. The physicist Seth Lloyd points out that any region of matter at all can be regarded as a quantum computer in which the bits are stored as particle spins, and the operations consist of the particles interacting with one another. Lloyd says that in some sense a kilogram-sized piece of ordinary matter is running a computation in which it updates a memory of 1031 bits at a rate of 1020 updates per second. And if the matter happens to be a black hole, the figures switch to a memory of 1016 bits being updated at a rate of 1035 updates per second. In either case, if we multiply the two numbers together, we get 1051 bit updates per second, which is something like a sextillion nonillion instructions per second, or a zetta-wattaflop. Lloyd suggests that a black hole computer could form for the tiniest fraction of a second, absorb computational input as mass and energy, carry out its vast computation, and dissolve into a pulse of energy containing the output. Dzeent! “I just had an idea!”97
When prognosticating, it’s easy to make the mistake of expecting future technology to be simply an amplified version of today’s. If the mechanical computer designer Charles Babbage were whisked to our time from the Victorian era, he might initially suppose that the buzzing beige box beneath my desk is stuffed with tiny clockwork gears. And we’d probably be just as mistaken to expect the PCs of the year 2500 to be using silicon chips. We may well discover new computing technologies that make the mind recipe feasible after all.
Another possibility regarding the creation of software to emulate the human brain is that there could be better approaches than the brute-force evolutionary search laid out in the mind recipe. In section 3.2: The Morphogenesis of a Brindle Cow I discussed the fact that many of the physical shapes found in living organisms are in fact patterns naturally formed by computations such as reaction-diffusion rules. The branching structure of a human hand isn’t so much something that was evolved in detail as it is a pattern that emerges from a type of reaction that takes place in an embryo. Now it may also be that, with a bit more insight, we’ll come to see that much of the brain’s organizational structure emerges naturally from certain rather simple kinds of morphogenetic processes. In this case, artificially growing a network similar to the brain might be radically easier than we’d supposed. So maybe we’ll get our intelligent robots fairly soon after all.
On the theme of computational futures, there’s an interesting idea first proposed by the science-fiction writer and computer-science professor Vernor Vinge in a 1993 talk.98 Vinge pointed out that if we can make technological devices as intelligent as ourselves, then there seems to be no reason that these devices couldn’t readily be made to run a bit faster and have a bit more memory so as to become more intelligent than people. And then—the real kicker—these superhuman machines might set to work designing still better machines, setting off a chain reaction of ever-more-powerful devices.
Vinge termed the potential event the Singularity. Although Vinge’s analysis is sober and scientific, in the last couple of decades, belief in his Singularity has become something of a cult among certain techies. Science-fiction writers, who have a somewhat more jaded view of predictions, have a saying about the enthusiasts: “The Singularity is the Rapture for geeks.” That is, among its adherents, belief in the Singularity has something of the flavor of the evangelical Christian belief in a world-ending apocalypse, when God will supposedly elevate the saved to heaven, leaving the rest of us to fight a final battle of Armageddon.
At one level, belief in the Singularity is indeed an instance of people’s age-old tendency to predict the end of the world. Once we have the Singularity, the machines can copy our brains and make us immortal. But once we have the Singularity, the machines may declare war on humanity and seek to exterminate us. Once we have the Singularity, the machines will learn how to convert matter into different forms and nobody will ever have to work again. But once we have the Singularity, the machines may store us in pods and use us as components. Once we have the Singularity, the machines will figure out how to travel faster than light and into the past. But once we have the Singularity, the machines will screw things up and bring the entire universe to an end. And so on.
Vinge describes several kinds of scenarios that could lead to a Singularity of cascading superhuman intelligence. We can group these somewhat science-fictional possibilities into three bins.
• Artificial minds. We design or evolve computing devices as intelligent as ourselves, and these entities continue the process to create further devices that are smarter than us. These superhuman computing devices might be traditional silicon-chip computers, nanotechnological assemblages, quantum computers, or bioengineered artificial organisms.
• Cyborgs. Humans split off a new species, part natural and part engineered. This could result either from bioengineering the human genome or from giving people an effortless, transparent interface to supercomputing helper devices. The resulting cyborgs will advance to superhuman levels.
• Hive minds. The planetary network of computers wakes up and becomes a superhuman mind. Alternately, people are equipped with built-in communication devices that allow society to develop a true group mind of superhuman powers.
I’ve already said enough about the artificial minds scenario, so let’s close this section with a few words about the other two.
The cyborg possibilities provoke what bioethicists call the “yuck factor.” Quite reasonably, we don’t like the idea of changing human biology. Gaia knows best. Don’t fix it if it ain’t broke. Keep the genie in the bottle!
But if we could become cyborgs via high-quality interfaces to external and detachable computing elements, it might not be so bad. In my science-fiction novels I often write about an uvvy (rhymes with “lovey-dovey”), which is meant to be a smart plastic communications-and-computation device that a person wears on the nape of the neck. For me an important aesthetic feature of the uvvy is that its plastic is soft, flexible, and flickering.
To keep down the yuckiness, a uvvy communicates with the user’s brain via magnetic fields rather than by poking hair-fine tendrils into the spine. An uvvy becomes something like a symbiotic partner, a bit like a really, really good cell phone.
One aspect of the cyborg scenarios is that they reduce the dichotomy between humans and machines. Depending on how you think about it, this can seem either good or bad. With a positive spin, the machines become our symbiotic partners and we advance together. With a negative spin, we see humanity being debased to the level of kludgy science experiments.
The hive mind scenarios represent a whole different way of thinking about computation—and this will be a topic I discuss in chapter five: The Human Hive.
Coming back to the starting point of this section, do I think that we’ll ever be able to make living mental copies of ourselves? It seems within the realm of possibility. But, in the end, people might feel it was too much trouble. After all, there’s no particular reason that any of us should be immortal. Nature’s perfectly happy to just keep growing fresh new people. Last year’s rose blossoms are gone, and it makes the world more interesting to have this year’s blooms be different from any that came before. Accepting my mortality gives me all the more reason to make the most of the time that I actually have.
• See book's home page for info and buy links.
• Go to web version's table of contents.
4.7: What Do You Want?
Consider the following bit of dialectical analysis.
• Universal automatism proposes a thesis: Your mental processes are a type of deterministic computation.
• Your sense of having a free will entails a seeming antithesis: Your thoughts and actions aren’t predictable.
• Wolfram advocates a beautifully simple synthesis: Your mind’s computation is both deterministic and unpredictable.
The synthesis is implicit in a conjecture that we’ve already mentioned several times.
• Principle of Computational Unpredictability (PCU). Most naturally occurring complex computations are unpredictable.
The workings of your mind are unpredictable in the sense that, when presented with a new input of some kind, you’re often quite unable to say in advance how you’ll respond to it.
Someone offers you a new job. Do you want it or not? Right off the bat, there’s no way to say. You have to think over the possibilities, mentally simulate various outcomes, and feel out your emotional responses to the proposed change.
Someone shows you a painting. Do you like it? Hold on. You have to think about the image, the colors, and the mental associations before you decide. Someone hands you a menu. What do you want to eat? Just a minute. You need to look into your current body feelings, your memories of other meals, your expectations about this particular restaurant.
We say a computation is unpredictable if there is no exponentially faster shortcut for finding out in advance what the computation will do with arbitrary inputs. When faced with an unpredictable computation, the only reliable way to find out what the computation does with some input is to go ahead and start up the computation and watch the states that it steps through.
Figure 105 shows the spacetime evolution of an unpredictable one-dimensional CA, with time running down the page. Even if I have full knowledge of this CA’s underlying rule and of the input pattern in a given row, the rule is gnarly enough that, in all likelihood, the only way I can possibly figure out the contents of a later row is to compute all the rows in between.
Certainly there’s no doubt that the endlessly flexible human mind embodies a universal computation, that is, a computation capable of emulating any other computation. Being universal, the human mind is class four and gnarly. Given this, we certainly expect the workings of the human mind to be unpredictable.
Once again, suppose I’m presented with some new input. Since my thoughts are unpredictable, the only way to find out what I’m going to end up thinking about the input is to go ahead and think until my mind is made up. And this means that, although my conclusion is in fact predetermined by how my mind works, neither I nor anyone else has any way of predicting what my conclusion will be. Therefore my thought process feels like free will.
But the process is, after all, deterministic, and deep down we all know this.
Of course, discerning the reasons behind our decisions can be hard. On the one hand, we can be unaware of our true motivations or unwilling to own up to them. “I have no idea why I did that.” And on the other hand, we sometimes like to suggest nobler reasons for our actions than the motivations more likely to have played a part. “I did it for his own good.” And, most crucial of all, even if we are blessed with full self-knowledge, it may be impossible to predict in advance how the influences of our various competing motivations will play out.
Figure 105: The Unpredictable China CA
This image follows a one-dimensional CA through twenty-four hundred generations, with six hundred generations per strip; the top of each strip is a continuation of the bottom of the strip to its immediate left. The world of this CA is 128 cells wrapped into a circle, meaning that the right and left edges of each strip match as well. If we were to paste everything together, this picture would be a cylinder like a baton. Notice the characteristic feature of class four rules: they send information back and forth by means of the moving patterns we call gliders. I found this rule in 1990 after about fifteen minutes of a Blind Watchmaker-style directed search, using a program that took me a year to write. I call it China because it looks a little like a silk fabric design. When a friend or loved one makes a decision, you ask why—and you expect an answer. Normally people do have reasons for doing things. And if they have reasons, then their decisions are in fact deterministic.
Really big decisions are rather rare in one’s life. Looking back on the few I’ve made, I wouldn’t say that any of them was capricious. Surprising to others, yes, but always logical for me. In each case, the course of action that I took was, for me at the time, the inevitable thing to do.
Two months ago, for instance, I decided to retire from teaching computer science at San Jose State University. I was tired of preparing new lectures and demos on difficult material, tired of wrestling with the ever-changing hardware and software, and eager to devote more time to my writing. These reasons might not have been enough, but then the state university faculty got an offer of a small golden handshake from the Terminator himself—that is, from the actor who once portrayed an implacable robot and who had recently become the governor of California. I was in a mental state where this financial offer was enough to tip the scales—and I went for it.
Where’s the free will in that? All I did was evaluate data, simulate alternate futures, and study my feelings. Given my mind-set and the various inputs, my retirement was inevitable. But if, immediately after the golden handshake offer, you’d asked me if I was going to retire, I wouldn’t have been able to give you a firm yes or no answer. I didn’t know yet. My deterministic computation wasn’t done, and for me it was unpredictable.
Just to drive home the point, let me quote the relevant passage from Wolfram’s A New Kind Of Science. (I’ve inserted some bracketed phrases to remind you that I’m using the word unpredictable as a synonym for what Wolfram prefers to call “irreducible.”)
Ever since antiquity it has been a great mystery how the universe can follow definite laws while we as humans still often manage to make decisions about how to act in ways that seem quite free of obvious laws.
But from the discoveries in this book it finally now seems possible to give an explanation for this. And the key, I believe, is the phenomenon of computational irreducibility [or unpredictability]...
For if the evolution of the system corresponds to an irreducible [or unpredictable] computation, then this means that the only way to work out how the system will behave is essentially to perform this computation—with the result that there can fundamentally be no laws that allow one to work out the behavior more directly.99
I’m quite happy with this resolution of the conflict between determinism and free will. But I find that when I try to explain it to people who aren’t universal automatists, they’re dubious. I’ll respond to three common objections.
Objection: “If I’m really like a computer program, then my free will is only an illusion. And I don’t want that to be true.”
Response: In the philosophical style, I’ll answer this with a counter question. By “free will” do you mean ability to make an utterly random decision? But what is “utterly random”? If something’s unpredictable, it’s all but indistinguishable from being random, no?
Some philosophers of science have tried to resolve the free will question by supposing that the brain has an ability to tap into a physically random process such as a chaotic system with unknown initial conditions, or a quantum measurement of some kind.
A universal automatist would reject this approach for two reasons. First of all, you don’t need to turn to physics since there are lots of dirt-simple rules that, when run upon a neural network like the brain, will generate unpredictable and random-looking data. And second, assuming that there’s a deterministic computation underlying the seeming uncertainties of quantum mechanics, all physical processes are deterministic, so you aren’t going to be able to get true randomness from nature either. Assuming that science will find its way past the wifty obfuscations of quantum mechanics, whatever seeming randomness you find in physics is just another example of the unpredictability of complex deterministic computations. So, coming back to the first point in this paragraph, you might as well accept that your mind is a deterministic computation.
Objection: “I can prove that I have free will by flipping a coin to make up my mind.”
Response: Even if our actions are deterministic, they are indeed influenced by the inputs that we get. The external world’s computations are something quite distinct from our own computations. Now it’s certainly possible that your deterministic thought process might tell you to flip a coin and to make a choice on the basis of what the coin says. But in this case, your actions still aren’t truly random. You’ve only added an extra coin-flip bit of input.
Objection: “My free will isn’t an illusion. I can prove it by doing the opposite of what I want to do.”
Response: That’s a contradiction. Once the dust settles, what you did is what you wanted to do. And you don’t really have free will over what you want to do—at least not in the sense of being off in some little control room and sending out command signals. Your drives and desires are influenced by biochemical cycles, memories, life experiences, and external inputs. You can make unexpected changes, but these are the results of your ever-flowing mental computation.
It’s valuable to realize that everyone’s mind is performing a gnarly class-four computation. Sometimes if I look at strangers, I’ll unkindly jump to the conclusion that they’re mindless robots—particularly if they don’t resemble me. Remember how your parents seemed to you when you were a teenager? Robots for sure.
One of the pleasant side effects of unexpected social interactions is that you get flashes of insight into the minds of people whom you might otherwise never meet. When I relax, I discover unexpected intricacies of emotion and humor within strangers. Nobody is simple on the inside. Simplicity is an impossibility. Every brain is carrying out a class four computation.
And this is no surprise, really. Look inward at your flow of thought. It’s like that China cellular automaton rule depicted in Figure 105. One thing leads to another. The gliderlike thought trains collide and light up fresh associations. Even if you’re lying in bed with your eyes closed, the flow continues, the endless torrent. Now and then you get stuck in a loop, but some unexpected glider eventually crashes in to break things up. You’re surfing the brain waves; and you yourself are the surf.
At this point, I’d like to mention a touchy subject: God. Let me immediately say that I’m not out to advocate religion. If you want to keep things more neutral, think of “God” as a convenient and colorful synonym for “the cosmos.” Or think of the “God” word as convenient shorthand for “the unknown.”
My reason for mentioning God is that there’s a particular connection between God and free will that intrigues me: When in dire straits, people sometimes ask God to help them change their behavior. And, often enough to matter, this seems to help them get better. What might this mean?
I would say that becoming desperate enough to turn to God involves recognizing a current inability to alter one’s mental patterns and a desire to attempt some higher-level change. The plea expresses a longing to jump out of a loop, a desire to move from one attractor to the next, a wish to experience a chaotic bifurcation.
If the plea works, does that mean that the Great Author, the Ground of All Being, the Omnipresent-Omnipotent-Omniscient One has reached down to change the parameters of some suffering character’s mental computations? And, more to the point, does this destroy determinism?
Well, we can keep determinism if we allow for a less supernatural view of reform-by-supplication. We could simply say that asking God for help has an organic effect upon a person’s brain. In other words, expressing a desire to have a spiritual life might activate, let us say, certain brain centers that release endorphins, which in turn affect the threshold levels of one’s neurons. And these changes nudge the brain activities to a new strange attractor. A deterministic chaotic bifurcation occurs.
Do I really think it works like that? Well, to be truthful, I’ve always felt comfortable about reaching out for contact with the divine. The world is big and strange, and we have only the barest inkling about what lies beneath the surface.
But even in this less materialistic view, a person can still be deterministic. Asking God for help in achieving a chaotic bifurcation is really no different from asking a doctor for penicillin. You can’t will an infection away, and you can’t will yourself to abandon some deeply ingrained bad habit. But at a slightly higher level, you may be able to muster the will to get help. And this higher level is, after all, simply part of your brain’s ongoing deterministic computation.
For that matter, God, too, could be deterministic. In the context of the theory I suggested at the end of section 2.5: What Is Reality?, God could be a deterministic nonreversible class four paratime metaphysical cellular automaton.
But that sounds so dull. Better to say the cosmos is dancing with us all the time. And that God is in the blank spaces between our thoughts—like in those white regions of the picture of the China CA.
• See book's home page for info and buy links.
• Go to web version's table of contents.
4.8: Quantum Soul
It’s valuable to remember how really odd it is to be conscious. The miracle of your mental life is being created by a carpet of cells that have grown themselves into a mat. How can this be?
Walking in the woods, I see a footbridge and automatically form a model of it. I step onto the bridge and look down at the ripples in the stream, the foam, the tiny standing waves. I’m thinking these are like the mind. A woman walks past, cautious of the silver-haired man she sees. I think of Joseph Campbell, of myth, of a fairy tale about a troll beneath a bridge. All this is coming from the meat weave in my head. The old associations are somehow alive in the background, always ready to pulse at my call, and new associations form as spontaneously as the eddies in the torrent below.
Yes, both humans and PCs are universal computers, so, in principle, each should be able to simulate the other. And comparing them sheds some light on both. But, no, I don’t think they’re very similar. You can build a cathedral from gray Lego blocks, but that’s not what the Notre-Dame is.
Figure 106: The Author with Nick Herbert at an April Fool’s Day Parade
One year I attended this annual parade in Boulder Creek, California.
Introspection makes me doubt the notion that the human mind works like a desktop computer. I’m abetted in this doubt by my friend Nick Herbert, one of the more colorful characters I’ve met in Silicon Valley. Nick started as a physicist designing hard drives, but these days is more likely to be found holding forth on consciousness. Here’s a quote from a thought-provoking article by him called “Quantum Tantra.”
By the high standards of explanation we have come to demand in physics and other sciences, we do not even possess a bad theory of consciousness, let alone a good one.
Speculations concerning the origin of inner experience in humans and other beings have been few, vague and superficial. They include the notion that mind is an “emergent property” of active neuronal nets, or that mind is the “software” that manages the brain’s unconscious “hardware.”...
Half-baked attempts to explain consciousness, such as mind-as-software or mind-as-emergent-property do not take themselves seriously enough to confront the experimental facts, our most intimate data base, namely how mind itself feels from the inside.100
Nick proposes that we think of the human mind as a quantum system. Recall that quantum systems are said to change in two ways: When left alone, they undergo a continuous, deterministic transformation through a series of
blended, “superposed” states, but when observed, they undergo abrupt probabilistic transitions into unambiguous “pure” states. Nick suggests that we can notice these two kinds of processes in our own minds.
• The continuous evolution of superposed states corresponds to the transcendent sensation of being merged with the world, or, putting it less portentously, to the everyday activity of being alert without consciously thinking much of anything. In this mode you aren’t deliberately watching or evaluating your thoughts.
• The abrupt transition from superposed state to pure state can be seen as the act of adopting a specific opinion or plan. Each type of question or measurement of mental state enforces a choice among the question’s own implicit set of possible answers. Even beginning to consider a question initiates a delimiting process.
The superposed states of quantum mechanics don’t fit in with classical physics, but, at an internal psychological level, superposed mental states are something we’re familiar with.
Note that computer scientists do have ways to model vagueness. For instance, neural nets provide outputs that can take on values intermediate between the definite yes-no values of one or zero. The notion of a continuous range of truth values is also studied as “fuzzy logic.”
But the blended, overlaid, superposed states of quantum mechanics really aren’t captured by intermediate truth values. The physicist Erwin Schrodinger once remarked that there’s a difference between a blurred photograph of a mountain and a crisp photo of a cloud. Being blurry is like having an intermediate truth value, and a cloud is like being in a superposed state that’s a blend of several pure states.
Let’s go back to the notion of coherence I discussed in section 2.6: How Robots Get High. Recall that a coherent system is one that independently evolves through a deterministic sequence of superposed states. And a decoherent system can become entangled with another system, such as a measuring device, that may even force it into a pure classical state. If we think of the mind’s two modes as the coherent and the decoherent mode, then it seems as if being asked a question moves a person from coherence toward decoherence. As I already mentioned, this usage is a little counterintuitive, in that the more someone talks about their opinions, the less quantum-mechanically coherent they become. But I’ve grown used to this terminology. Indeed, as a kind of extended thought experiment, I had the villains in my recent epic novel Frek and the Elixir decohere their victims by interrogation.
Jayney leaned over Frek, her plastic dragon face bent into a parody of a motherly smile. She had fangs. She bit his neck and drew something out of him, leaving numbness in its place.
A hundred bland doughy faces seemed to bloom around him, pressing forward, staring, asking questions—all at the same time. Willy-nilly, Frek was unable to do anything but answer.
“How old are you? Are you happy? What’s your name? Do you miss home? How tall are you? Are you frightened?”
With each response, Frek became a bit less himself and more of a statistic. The questions were flattening him out. On and on they came.
“Rate your feelings about the following on a scale from one to five, ranging from dislike very much to like very much...
“Rate your perceived frequency of the following classes of events on a scale from one to five, ranging from almost never to extremely often...
“Specify your agreement with the following statements on a scale from one to five, ranging from disagree strongly to agree very much...”
The questions came at him thick and fast. In a few minutes, every spark of Frek’s own true self had been sapped away. All he felt now was a faint ache all through his bones, like the pain from a bad tooth.
Frek was a rag doll, an automaton, a thing. He contained no mysteries. He was fully decoherent.101
Isn’t that, really, a bit how it feels when someone starts firing questions at you? It’s unpleasant when someone substitutes interrogation for the natural flow of conversation. And it’s still more unpleasant when the grilling is for some mercenary cause. You have every reason to discard or ignore the surveys with which institutions pester you. (But please don’t use this reasoning as an excuse not to vote!)
Getting back to my main line of thought, let’s see how Nick Herbert’s notion of two kinds of mental processes fits in with the dialectic triad I laid down at the start of chapter one: Computation Everywhere. This time I’ll list two possible kinds of synthesis between the lifebox and the soul. These will be the “gnarliness” synthesis that I’m primarily arguing for, and an alternate quantum mind synthesis based on Nick Herbert’s ideas.
• Thesis (Lifebox): Our theoretical knowledge of computational universality and our practical experience with neural nets and genetic algorithms suggests that any clearly described human behavior can eventually be emulated by a deterministic computation.
• Antithesis (Soul): Upon introspection we feel there is a residue that isn’t captured by any scientific system; we feel ourselves to be quite unlike machines. This is the sense of having a soul.
• Gnarliness Synthesis (Seashell): If you’re a complex class-four computation, you don’t feel like you’re predictable, even though in fact you are a fully deterministic process. Computational systems can generate beautiful and unexpected patterns. A complex computation could perfectly well become conscious and feel itself to have a soul.
• Quantum Mind Synthesis: The soul can be given a scientific meaning as one’s immediate perception of one’s coherent uncollapsed wave function, particularly as it is entangled with the uncollapsed universal wave function of the cosmos.
Let’s say a bit more about the quantum mind idea. How is it that the human brain could function in the two kinds of modes? Is there something specific about the human brain that allows us to couple our coherent superposed-state experiences with an ability to collapse down into discrete, pure states congenial to a PC?
The physicist Roger Penrose and the psychologist Stuart Hameroff point out that biological cells such as neurons include networks of very fine structures called microtubules. They suggest that the microtubles might somehow serve as a locus for quantum computations, and that it might be physically true that the brain at times is working in parallel superposed states. Most scientists dismiss this notion, citing problems with staying in a specific superposed quantum state for any appreciable length of time in a body-temperature assemblage like a brain. But there are methods of quantum error correction that might possibly make it possible for elements of the brain to be in coherent states for appreciable lengths of time.102
Whether or not Penrose’s somewhat unpopular ideas pan out, they’re at least an illustration of how a brain might include some truly quantum mechanical kinds of computation. Quoting the same essay by Nick Herbert:
Looking inside, I do not feel like “software,” whatever that might mean, but indeed like a shimmering (wavelike?) center of ambiguous potentia (possibilities?) around which more solid perceptions and ideas are continually congealing (quantum jumps?). This rough match of internal feeling with external description could be utterly deceptive but it at least shows that the quantum model of mind can successfully confront the introspective evidence in a way that no other mind models even attempt.
Even were our brains to exhibit large-scale quantum behavior, there’s no need to be human chauvinists. The quantum mind synthesis of the lifebox and the soul doesn’t rule out the possibility that machines or biocomputing devices could yet be equivalent to us. For any physical object is, after all, subject to quantum mechanics. Certain kinds of devices could well remain coherent and have uncollapsed wave functions for protracted periods of time.
As it so happens, building such devices is precisely what many quantum computing researchers are trying to do. My sense of Nick Herbert’s somewhat visionary ideas is that he is trying to imagine how quantum computation would feel from the inside—and discovering in the process that it’s something we do all the time.
In the fall of 2002, I was thinking about coherence a lot, largely in terms of the many-universes interpretation of quantum mechanics. I would occasionally reach the point where I was able to turn off my forever-talking inner narration and feel as if I had spread out into a quantum mechanical union with the world. One memorable December afternoon in Paris, I felt like I’d merged with the smoke from a smokestack. Here’s how I described it in my journal. (See Rudy Rucker, Journals 1990-2014, p. 463)
I keep working on this new mental exercise of becoming coherent, of being in a superposed state, of existing in multiple parallel universes, and that feels very good. Walking in the Latin Quarter, looking at some smoke from a chimney against the sky, not naming it, just seeing it, letting its motions move within my mind, I realize I’m no different than a computer screen showing a two-dimensional cellular automaton, with the smoke like a touch-cursor dragged across my brain. I am entangled with the smoke. I am coherent, but my coherence includes the smoke, I have joined the system, merged it into me. Like the old koan, Q: I see a flag is blowing in wind: is the flag moving or is the wind moving? A: My mind is moving. Finally I get it, a nice moment of aha, a satori in Paris.
One final point. The distinction between two kinds of quantum mechanical processes rests on the standard Copenhagen interpretation of quantum mechanics. In section 2.5: What Is Reality? I discussed John Cramer’s transactional interpretation of quantum mechanics, under which events really are determined and the superposed states are more in the nature of an illusion. The price Cramer pays for his model is in stipulating that the future influences the past or, putting it in a less time-bound fashion, his spacetime is an undivided whole, filled with patterns of synchronistic symmetries between the past and the future.
The notion of past and future being in harmony is strange to the usual view of physics, but it’s the daily bread of writers. If the first line and the second line of a poem both end in “oo,” does the first “oo” cause the second, or does the second “oo” cause the first? Neither, of course. Rhyming line endings are part of the poem’s pattern as a whole. When a hero’s death by fire is prefigured by a candle flame in a movie’s opening scene, the chain of causation really leads neither forward nor backward but—sideways, through paratime to the germ of the plot in the mind of the screenwriter.
For a universal automatist, the win in Cramer’s view is that it becomes possible to regard spacetime as resulting from a deterministic computation oriented along the second dimension of time that I call paratime. But do we then lose Nick Herbert’s rather attractive metaphor of the quantum mind?
Not entirely. We can still keep a notion of higher consciousness. But now, rather than regarding it as being the experience of a superposed state, we might instead view it as an experience of the world as a timeless whole, or perhaps as an experience of the world’s higher causation along the axis of paratime.
Words, words, words. I just stepped outside to take a break. The air is cool and fresh. I started early today, and the sun is still low. Dewdrops sparkle on the blades of grass, each drop holding its own idiosyncratic image of the world. I get my camera and take a picture (Figure 107).
In this chapter I’ve been describing how to view the mind as a deterministic computation. But at the end of the analysis, I still feel that something’s missing, some breath of human life.
Although I mentioned at the start of this chapter that the Egyptians thought of the heart as being of high importance, I’ve gone ahead and spent the whole chapter talking about the brain.
Figure 107: Sparkling Dew
So now, finally, here’s some heart to balance things out.
Twenty years ago, on my thirty-ninth birthday, my beautiful wife, Sylvia, wrote a lovely and lovable poem urging me to set aside my endless philosophizing and pay attention to her.
It’s your birthday!
Let down your proofs—
Count my numbers,
Process my words,
Weigh my mass,
And square my root!
Feel my fractals,
Join my space—
Would a robot ever write a poem like that? Well, maybe...someday. But not anytime soon. It’s important not to confuse philosophical dreams with the actual world we live in right now. Turn off the computer and give your partner a kiss. This means you, Rudy.