Part II: How Brains Might Work
Two weeks ago I wrote the first part of this column in which I made an attempt to explain how it is that we are able to design very complex machines like computers: we do it by employing a hierarchy of concepts, each layer of which builds upon the layer below it, ultimately allowing computers to perform seemingly miraculous tasks like beating Gary Kasparov at chess at the highest levels of the hierarchy, while all the way down at the lowest layers, the only thing going on is that some electrons are moving about on a tiny wafer of silicon according to simple physical rules. [Photo shows Kasparov in Game 2 of the match.] I also tried to explain what gives computers their programmable flexibility. (Did you know, for example, that Deep Blue, the computer which drove Kasparov to hair-pulling frustration and humiliation in chess, now takes reservations for United Airlines?)
But while there is a difference between understanding something that we ourselves have built (we know what the conceptual layers are because we designed them, one at a time, after all) and trying to understand something like the human brain, designed not by humans but by natural selection, there is also a similarity: brains also do seemingly miraculous things, like the writing of symphonies and sonnets, at the highest levels, while near the bottom we just have a bunch of neurons connected together, digitally firing (action potentials) away, again, according to fairly simple physical rules. (Neuron firings are digital because they either fire or they don’t–like a 0 or a 1–there is no such thing as half of a firing or a quarter of one.) And like computers, brains are also very flexible at the highest levels: though they were not designed by natural selection specifically to do so, they can learn to do long-division, drive cars, read the National Enquirer, write cookbooks, and even build and operate computers, in addition to a million other things. They can even turn “you” off, as if you were a battery operated toy, if they feel they are not getting enough oxygen, thereby making you collapse to the ground so that gravity can help feed them more of the oxygen-rich blood that they crave (you know this well, if you have ever fainted).
To understand how brains do all this, this time we must attempt to impose a conceptual framework on them from the outside, as it were; a kind of reverse-engineering. This is what neuroscience attempts to do, and as I promised last time, today I would like to present a recent and interesting attempt to construct just such a scaffolding of theory on which we might stand while trying to peer inside the brain. This particular model of how the brain works is due to Jeff Hawkins, the inventor of the Palm Pilot and the Treo Smartphone, and a well-respected neuroscientist. It was presented by him in detail in his excellent book On Intelligence, which I highly recommend. What follows here is really just a very simplified account of the book.
Let’s jump right into it then: Hawkins calls his model the “Memory-Prediction” framework, and its core idea is summed up by him in the following four sentences:
The brain uses vast amounts of memory to create a model of the world. Everything you know and have learned is stored in this model. The brain uses this memory-based model to make continuous predictions of future events. It is the ability to make predictions about the future that is the crux of intelligence. (On Intelligence, p. 6)
Hawkins focuses mainly on the neocortex, which is the part of the brain responsible for most higher level functions such as vision, hearing, mathematics, music, and language. The neocortex is so densely packed with neurons, that no one is exactly sure how many there are, though some neuroscientists estimate the number at about thirty billion. What is astonishing is to realize that:
Those thirty billions cells are you. They contain almost all your memories, knowledge, skills, and accumulated life experience… The warmth of a summer day and the dreams we have for a better world are somehow the creation of these cells… There is nothing else, no magic, no special sauce, only neurons and a dance of information… We need to understand what these thirty billion cells do and how they do it. Fortunately, the cortex is not just an amorphous blob of cells. We can take a deeper look at its structure for ideas about how it gives rise to the human mind. (Ibid., p. 43)
The neocortex is a thin sheet consisting of six layers which envelops the rest of the brain and is folded up in a crumpled way. This is what gives the brain its walnutty appearance. (If completely unfolded, it would be quite thin–only a couple of millimeters–and would cover an area about the size of a large dinner napkin.) Now, while the neocortex looks pretty much the same everywhere with its six layers, different regions of it are functionally specialized. For example, the Broca’s area handles the rules of linguistic grammar. Other areas of the neocortex have also been mapped out functionally in quite some detail by techniques such as looking at brains with localized damage (due to stroke or injury) and seeing what functions are lost in the patient. (Antonio Damasio presents many fascinating cases in his groundbreaking book Descartes’ Error.) But while everyone else was looking for differences in the various functional areas of the cortex, a very interesting observation was made by a neurophysiologist named Vernon Mountcastle (I was fortunate enough to attend a brilliant series of lectures by him on basic physiology while I was an undergraduate!) at Johns Hopkins University in 1978: he noticed that all the different regions of the neocortex look pretty much exactly the same, and have the same structure, whether they process language or handle touch. And he proposed that since they have the same structure, maybe they are all performing the same basic operation, and that maybe the neocortex uses the same computational tool to do everything. Mountcastle suggested that the only difference in the various areas are how they are connected to each other and to other parts of the nervous system. Now Hawkins says:
Scientists and engineers have for the most part been ignorant of, or have chosen to ignore, Mountcastle’s proposal. When they try to understand vision or make a computer that can “see,” they devise vocabulary and techniques specific to vision. They talk about edges, textures, and three-dimensional representations. If they want to understand spoken language, they build algorithms based on rules of grammar, syntax, and semantics. But if Mountcastle is correct, these approaches are not how the brain solves these problems, and are therefore likely to fail. If Mountcastle is correct, the algorithm of the cortex must be expressed independently of any particular function or sense. The brain uses the same process to see as to hear. The cortex does something universal that can be applied to any type of sensory or motor system. (Ibid., p. 51)
The rest of Hawkins’s project now becomes laying out in detail what this universal algorithm of the cortex is, how it functions in different functional areas, and how the brain implements it. First he tells us that the inputs to various areas of the brain are essentially similar and consist basically of spatial and temporal patterns. For example, the visual cortex receives a bundle of inputs from the optic nerve, which is connected to the retina in your eye. These inputs in raw form represent the image that is being projected onto the retina in terms of a spatial pattern of light frequencies and amplitudes, and how this image (pattern) is changing over time. Similarly the auditory nerves carry input from the ear in terms of a spatial pattern of sound frequencies and amplitudes which also varies with time, to the auditory areas of the cortex. The main point is that in the brain, input from different senses is treated the same way: as a spatio-temporal pattern. And it is upon these patterns that the cortical algorithm goes to work. This is why spoken and written language are perceived in a remarkably similar way, even though they are presented to us completely differently in simple sensory terms. (You almost hear the words “simple sensory terms” as you read them, don’t you?)
Now we get to one of Hawkins’s key ideas: unlike a computer (whether sequential or parallel), the brain does not compute solutions to problems; it retrieves them from memory: “The entire cortex is a memory system. It isn’t a computer at all.” (Ibid., p. 68) To illustrate what he means by this, Hawkins provides an example: imagine, he says, catching a ball thrown at you. If a computer were to try to do this, it would attempt to estimate its initial trajectory and speed and then use some equations to calculate its path, how long it will take to reach you, etc. This is not anything like what your brain does. So how does your brain do it?
When a ball is thrown, three things happen. First, the appropriate memory is automatically recalled by the sight of the ball. Second, the memory actually recalls a temporal sequence of muscle commands. And third, the retrieved memory is adjusted as it is recalled to accomodate the particulars of the moment, such as the ball’s actual path and the position of your body. The memory of how to catch a ball was not programmed into your brain; it was learned over years of repetitive practice, and it is stored, not calculated, in your neurons. (Ibid., p. 69)
At first blush it may seem that Hawkins is getting away with some kind of sleight of hand here. What does he mean that the memories are just retrieved and adjusted for the particulars of the situation? Wouldn’t that mean that you would need millions of memories for every single scenario like catching a ball, because every situation of ball-catching can vary from another in a million little ways? Well, no. Hawkins now introduces a way of getting around this problem, and it is called invariant representation, which we will get to soon. Cortical memories are different from computer memory in four ways, Hawkins tells us:
- The neocortex stores sequences of patterns.
- The neocortex recalls patterns auto-associatively.
- The neocortex stores patterns in an invariant form.
- The neocortex stores patterns in a hierarchy.
Let’s go through these one at a time. The first feature is why when you are telling a story about something that happened to you, you must go in sequence (and why often people include boring details in their stories!) or you may not remember what happened; like only being able to remember a song if you sing it to yourself in sequence, one note at a time. (You couldn’t recite the notes backward–or even the alphabet backward very fast–while a computer could.) Even very low-level sensory memories work this way: the feel of velvet as you run your hand over it is just the pattern of very quick sequential nerve firings that occurs as your fingers run over the fibers. This pattern is a different sequence in case you are running your hand over gravel, say, and that is how you recognize it. Computers can be made to store memories sequentially, such as a song, but they do not do this automatically, the way the cortex does.
Auto-associativity is the second feature of cortical memory and what it means is that patterns are associated with themselves. This makes it possible to retrieve a whole pattern when only a part of it is presented to the system.
…imagine you see a person waiting for a bus but can only see part of her because she is standing partially behind a bush. Your brain is not confused. Your eyes only see parts of a body, but your brain fills in the rest, creating a perception of a whole person that’s so strong you may not even realize you’re only inferring. (Ibid., p. 74)
Temporal patterns are also similarly retrieved and completed. In a noisy environment we often don’t hear every single word that someone is saying to us, but our brain fills in with what it expects to have heard. (If Robin calls me on Sunday night on his terrible cell phone and says, “Did you …crackle-pop… your Monday column yet?” My brain will automatically fill in the word “write.”) Sequences of memory patterns recalled auto-associatively essentially constitute thought.
Now we get to invariant representations, the third feature of cortical memory. Notice that while computer memories are designed for 100% fidelity (every bit of every byte is reproduced flawlessly), our brains do not store information this way. Instead, they abstract out important relationships in the world and store those, leaving out most of the details. Imagine talking to a friend who is sitting right in front of you. As you talk to her, the exact pattern of pixels coming over the optic nerve from your retina to your visual cortex is never the same from one moment to another. In fact, if you sat there for hours, no pattern would ever repeat because both of you are moving slightly, the light is changing, etc. Nevertheless you have a continuous sense of your friend’s face being in front of you. How does that happen? Because your brain’s internal pattern of representation of your friend’s face does not change, even though the raw sensory information coming in over the optic nerve is always changing. That’s invariant representation. And it is implemented in the brain using a hierarchy of processing. Just to give a taste of what that means, every time your friend’s face or your eyes move, a new pattern comes over the optic nerve. In the visual input area of your cortex, called V1, the pattern of activity is also different each time anything in your visual field moves, but several levels up in the hierarchy of the visual system, in your facial recognition area, there are neurons which remain active as long as your friend’s face is in your visual field, at any angle, in any light, and no matter what makeup she’s wearing. And this type of invariant representation is not limited to the visual system but is a property of every sensory and cortical system. So how is this invariant representation accomplished?
I’m sorry, but unfortunately, I have once again run out of time and space and must continue this column next time. Despite my attempts at presenting Hawkins’s theory as concisely as possible, it is not possible to condense it further without losing essential parts of it and there’s still quite a bit left, and so I must (reluctantly) write a Part III to this column in which I will present Hawkins’s account of how invariant representations are implemented, how memories are used to make predictions (the essence of intelligence), and how all this is implemented in hierarchical layers in the actual cortex of the brain. Look for it on May 8th. Happy Monday, and have a good week!