On Intelligence by Jeff Hawkins with Sandra Blakeslee
by Jeff Hawkins with Sandra Blakeslee
Times Books. 272 pp. $25.00
Elementary biology teaches us that many of the vital parts of the body have forms that follow their function. The heart has tubes and valves like a hydraulic pump; the lungs expand and deflate like bellows; the ear has an intricate apparatus of bones and membranes that transmit sound like a stick beating on a drum. The fact that each of these organs can be understood in reasonably simple physical terms has enabled us to construct clever mechanical prostheses—like artificial hearts, respirators, and cochlear implants—that do their jobs, more or less well, when things go wrong.
It goes without saying that this is not true for the brain, which resembles nothing so much as a mass of pinkish goo. Aristotle, reasoning from inspection, asserted that the brain was a reservoir for cooling the blood; as late as the 1890′s, most biologists believed that it had roughly the structure of a sponge. Only at the turn of the last century did the great anatomist Santiago Ramón y Cajal begin to unlock the secrets of the brain’s architecture, describing in exquisite detail the ways in which its cells, called neurons, articulate with each other by sending out a myriad of protoplasmic wires (dendrites) and cables (axons).
Cajal noted, moreover, that in the outermost and largest part of the human brain—the tortuously furrowed cerebral cortex—the neurons appear to be arranged in six layers, with axons from each layer projecting to neurons in other layers in a stereotypical fashion. Undoubtedly this observation was momentous: since the cortex is almost certainly the seat of intelligence (it is large in humans, small in rats, and almost nonexistent in fish), it follows that understanding its construction is key to understanding the mind. Such understanding, however, remained beyond Cajal’s grasp. With the allegorical flair of a Spanish mystic, he likened the beautiful and enigmatic neurons of the cortex to “butterflies of the soul.”
Since Cajal’s time, biologists have learned a great deal about the physiology of neurons. We know how they are born, migrate to their ultimate sites within the brain, form connections, and communicate with other neurons by releasing packets of chemicals that produce electrical signals. What we have not learned, however, is the purpose served by the six-layered organization of the cerebral cortex. Put simply, the matter of how the structure of the brain produces thought remains almost totally mysterious.
Enter Jeff Hawkins. A computer engineer by training, Hawkins is best known as the inventor of the PalmPilot and other handheld gadgets. But his true intellectual passion lies elsewhere: as he tells it, he has searched all his life for a comprehensive and intuitive explanation of how the brain works. Nor does he make any secret of his goal: coming from an engineering perspective, Hawkins believes that a working model of the cortex can serve as a blueprint for truly intelligent machines, capable of solving abstract problems, testing theories, and making reliable predictions about everything from the weather to political unrest.
On Intelligence, written with the help of the veteran New York Times science writer Sandra Blakeslee, is Hawkins’s attempt to lay out just such a model, and to do so “in a way that anybody will be able to understand.” True to this aim, On Intelligence is written in simple, straightforward prose, with short sentences and few rhetorical flourishes. There is some unavoidable technical jargon, but it is kept to a minimum. Hawkins provides numerous examples and thought experiments along the way, encouraging the reader to discern the workings of the cortex in experiences as diverse as appreciating a melody and drying oneself with a towel.
Hawkins’s basic insight, borrowed from the neurobiologist Vernon Mountcastle, is that if the structure of the cortex is basically uniform, every part of it must be performing a similar function. To be sure, the brain processes different kinds of information in different parts of the cortex: visual sensations feed into an area at the back of the head, while motor commands are sent out by neurons residing in a strip that lies just in front of the ears. But Hawkins’s idea is that all of this information is processed in basically the same way. No matter what the sensory input, the brain is essentially receiving patterns of stimulation with spatial and temporal properties.
When we look out at the world, we perceive a display of light that changes over time; when we touch something, we feel a texture that changes as we move our fingers. Because these patterns have certain regularities—the parts of an object tend to move together; eyes tend to occur with noses, and so forth—the cortex can use them to build a model of the world. This, according to Hawkins, is memory, and it has a number of important properties.
First, the cortex stores memories in the form of sequences of patterns: the succession of notes in a song, the eye-nose-eye sequence of a face, even the sequence of towel rubs and pats we use to dry off. Second, these memories are self-reinforcing, or “auto-associative”: as soon as one part of a pattern is activated, the entire pattern follows. Hum the first bar of a song, and the rest comes to mind automatically. Third, memories are, at some level, stored in a way that is independent of irrelevant variations; we recognize a melody regardless of pitch, and a face regardless of the angle from which we see it.
On Intelligence proposes that these properties of memory allow us to make predictions based on the information that has accumulated in our cortices. When you climb into a rental car you have never seen before, you nevertheless recognize its various parts, and can predict that if you engage the sequence of actions appropriate to driving (turning the key and pressing the gas pedal), the car should move; if it does not, something is wrong with the world, or at least with the car. This kind of prediction-making based on memory is what Hawkins calls intelligence, and it constitutes what he believes is the primary function of the cortex.
Hawkins goes into some detail about the ways in which the six-layered structure of the cortex is particularly suited to storing sequences of auto-associative memories. The essence of his proposal is that the cortex is arranged in a hierarchy: parts of the brain that are responsible for representing very small bits of patterns (say, particular sounds) send specific projections to higher-up areas that represent progressively larger and more abstract bits (notes, intervals, bars, songs). The higher-up areas in turn send feedback to lower areas, instructing them what to expect next and monitoring unexpected events. At the top of the hierarchy are those areas that integrate and coordinate memories of different types, checking the sum of what we perceive against the model of reality we have built by experience.
“Predictability is the very definition of reality,” writes Hawkins. Therefore, because the cortex can predict, it is capable of apprehending the world. Creativity, in this view, is nothing more than the ability to make uncommon predictions by analogy—likening one pattern to another (“love is a smoke made with the fume of sighs”; “life is like a box of chocolates”) and deriving predictions from it. Imagination is nothing more than the generation of predictions that feed into other predictions.
There is no question that this memory-prediction model is simplistic. Hawkins cheerfully admits that he has ignored or glossed over reams of neuroscientific data—both for the sake of accessibility to the lay reader and because he believes that rendering the subtleties in the scientific literature would only serve to obscure the big picture he wants to paint. The question is, however, whether this big picture is basically accurate and just needs to be fleshed out in the details, or whether it is so vague and underspecified that it provides no useful outline at all.
The framework laid out in On Intelligence has considerable intuitive appeal. As Hawkins illustrates with numerous examples, one of the general properties of intelligence is the ability to make sense of the world, and to make predictions about the present and the future based on what we already know. The mind is particularly adept at seeing patterns, and can even be tricked into evoking them based on the scantiest of sensory data; this is why we see faces in Martian rocks and in oddly shaped corn flakes. Indeed, it is probably why we dream. Random firing of neurons in the visual part of the cortex may activate part of a visual pattern; this pattern is automatically completed by other parts of the cortex, and that in turn triggers the retrieval of associated sights, sounds, and memories. The filtering of such activity up and down the cortical hierarchy creates a seamless perceptual experience that has no basis in reality, but nevertheless conforms to our knowledge about the structure of the world.
Even on a sympathetic reading, however, Hawkins’s treatment of the facts is disturbingly cavalier. He adopts some rather unconventional hypotheses, not because they are particularly well-supported by scientific fact but because they fit with his theory. For example, because his argument stipulates that memories are formed in the cortex, he embraces the idea that the hippocampus, generally considered to be an ancient part of the brain involved in memory formation, is actually a newly evolved structure that registers violations of expectancy.
At the same time, Hawkins’s appreciation of cognition seems in many instances to be grossly unsophisticated. He claims that “when we assign a name to something, we do so because a set of features consistently travels together.” Anyone who has taken a freshman course in philosophy knows that this kind of assertion is liable to collapse as soon as we try to identify the definitional features of the “something,” be it a “game” or a “bachelor,” or to explain why a cat that has been doused in musk and painted to look exactly like a skunk is nevertheless still very much a cat.
But the weakest link in Hawkins’s theory is also its main selling point: namely, his assumption that the cortex is an all-purpose learning device capable of solving virtually any problem by making predictions based on prior experience. In fact, a preponderance of evidence suggests that much of what we know—or are capable of learning—is not determined by experience, but is hardwired into the brain before birth. Indeed, it might logically be impossible to learn many of the things that we learn if we did not enter the world without some innate assumptions about its structure: we would simply never get enough input to make the correct predictions.
A prime example is language. Hawkins rather glibly asserts that we are born without knowledge of language, just as we were born without knowledge of houses; he assumes that “the cortex has a clever learning algorithm” that naturally extracts grammar from exposure to speech, without any dedicated language machinery.
Everything we know about language and the brain suggests that this is false. Children appear to be born with brains genetically rigged to acquire human language; monkeys, despite their well-developed cortices and the heroic efforts of many a psychologist, never seem to get much more than the capacity to ask for a banana. Even otherwise intelligent humans do not learn language if they are not exposed to it in the critical period before puberty. In this light, Hawkins’s prediction that his wonderfully plastic cortical algorithm can one day be adapted to produce computers capable of “intelligent speech recognition” seems wildly off the mark—unless, perhaps, one is willing to ascribe intelligence to the linguistic prowess of a gray parrot.
And this gets to the heart of the problem with On Intelligence. The cortex as Hawkins envisions it is no doubt capable of solving many problems, and it is probably reasonable to assume that the principles he sets forth can be applied in order to engineer things like “smart” cars that can navigate obstacles, and “smart” cameras that recognize objects from different angles. But it is worth keeping in mind that rats, too, can do these things. In this respect, both rats and probably even lobsters may be more intelligent than contemporary computers. Nevertheless, there remains an important qualitative distinction between this kind of intelligence and the intelligence possessed by human beings.
Hawkins’s theory leaves no room for such a distinction. On his account, the primary difference between the human cortex and the rat cortex is that the human cortex is much bigger. This is true, of course, but it is also true that the human cortex is differently wired, and comes equipped with a set of specialized functions—language, music, and numerical reasoning, to name a few. Most likely, these are not solely the product of greater computing power; if they were, we would presumably be able to chart flawless courses over thousands of miles like migrating geese, and remember the locations of hundreds of objects like squirrels who stash acorns in the winter.
Instead, we take wrong turns and lose our keys, but we can speak and write books about intelligence. This is the unique endowment of human beings. It is implemented, somehow, in the cortex, but in the end, Hawkins does not get us much closer to understanding how.
In one of the numerous and distracting autobiographical interludes that punctuate On Intelligence, Hawkins informs us that in 1986 he left a lucrative position in Silicon Valley to enroll as a graduate student in biophysics at the University of California at Berkeley. He does not tell us what later impelled him to leave academia and return to the entrepreneurial world, but one imagines he realized that the grindstone of research was getting him no closer to his goal of a grand theory of the brain.
The younger Hawkins might have benefited by consulting the advice of Cajal, who in 1916 wrote with penetrating disdain about brilliant thinkers who preferred constructing audacious theories to discovering facts by precise observation. “The essential thing for them is the beauty of the concept,” he observed. “It matters very little whether the concept itself is based on thin air, so long as it is beautiful and ingenious, well-thought-out and symmetrical.” Hawkins’s theory is all of these things. And that may be precisely what is wrong with it.