On Intelligence , by Jeff Hawkins -- Book Review

Book Review

Computer expert Jeff Hawkins, inventor of the PalmPilot and the method of using a stylus to enter text, "fell in love with brains" [233] in 1979. He thereupon went to U.C. Berkeley to study neurology and biophysics. Initially convinced that "solving the puzzle of intelligence was something that could be achieved in my lifetime," he now asserts that "we may be able to create useful prototypes and cortical simulations within just a few years." [233]

That's right, machines with true intelligence, made of silicon instead of living cells.

Functionalism is the belief that intelligence is based on the organization of a system's parts, and not dependent on whether the system is made out of neurons, computer chips, or something else. The author describes himself as a functionalist. In fact, he says, nature's biological solution is somewhat of a Rube Goldberg contraption. Just as airplanes aren't built like birds, other forms of hardware may be used to make intelligent machines. [36-38]

You've heard of artificial intelligence (AI), but that is not what Hawkins is talking about. He doesn't mean mechanisms which use very complicated internal processes to simulate human behaviors, such as computers which can playing chess. He's talking about mechanisms which, like the neurons of the human cortex, derive wonderous results by repeating a single common algorithm in billions of places.

About the Authors

Jeff Hawkins is one of the most successful and highly regarded computer architects and entrepreneurs in Silicon Valley. He founded Palm Computing and Handspring, and created the Redwood Neuroscience Institute to promote research on memory and cognition. Also a member of the scientific board of Cold Spring Harbor Laboratories, he lives in northern California.

Sandra Blakeslee has been writing about science and medicine for The New York Times for more than thirty years and is the co-author of Phantoms in the Brain by V. S. Ramachandran and of Judith Wallerstein's bestselling books on psychology and marriage. She lives in Santa Fe, New Mexico.

- From the Publisher

Imaging of the brain [31, 52-53] may have contributed to the mistaken assumption that different parts of the brain perform their respective tasks in entirely different ways. Attention to specialized locations has distracted us from noticing the time-varying pattern, the common algorithm used everywhere.

The author rejects the idea that the brain is little understood because it is "so complicated." Instead, he believes, "a few intuitive but incorrect assumptions" have prevented understanding of a process that is simpler than we had previously believed. [5-6]

In the 1970s, neuroscientist Vernon Mountcastle discovered that there is uniformity of the human cortex, regardless of whether we consider the sections that process vision, hearing, touch, etc. He hypothesized that somehow the different regions were performing the same tasks, despite the variety of results. The multitude of the mental tasks derive from "the same basic operation" and using "the same computational tools" [50] Researchers in recent decades have tended to ignore Mountcastle's discovery, although he had discovered what Hawkins considers "the Rosetta Stone of neuroscience." [51-52] As a result, we now have "a lot of data without a theory." [10]

A most basic premise of the book is that, contrary to the belief of mathematician Alan Turing, who contributed much to digital computer architecture, intelligence is not defined in terms of externally observed behavior. This is evident from that fact that "you can be intelligent just lying in the dark, thinking and understanding." [29] Instead, intelligence is defined by hierarchical [44-46] organization.

In human perception, lower cortical layers exhibit constantly-varying states, and higher layers form "invariant representations." [69 et. seq.] For example, with vision, the first neurons to receive signals from the optic nerve exhibit nothing but rapid on-and-off behavior, since each eye moves in jerky motions (saccades) [57, 110]. Further up the hierarchy in the visual cortex, certain "face cells" are in a steady state as long as a a certain friend's face is in the field of vision [77, 113]. Similarly with hearing, neurons in the primary auditory area discharge with high frequency, in response to sound, which is the variation in air pressure, thousands of times per second, on the eardrum. "But a probe stuck into a higher auditory regions should find cells that fire steadily every time 'Three Blind Mice' is being played, without concern for the instrument, tempo, or other details." [115] Providing experimental evidence for the idea that invariant representations are stored, there was a discovery of neurons in one subject the state of which indicated whether or not the person was currently looking at a picture of Bill Clinton. [108]

The neocortex, an evolutionary addition which sits upon the 'primitive brain', is found only in the brains of mammals. Furthermore, humans are unique among mammals because of the large area covered by the layers of cortex cells. [42, 98-99] Hawkins writes, "We see deeper analogies, more structure on structure, than other mammals." [180] The neocortex has approximately 30 billion neurons [43] and approximately 30 trillion synapses. [49]

An astonishing discovery is that the brain has more feedback than forward circuitry. "As you read aloud, higher regions of your cortex send more signals 'down' to your primary visual cortex than your eye receives from the printed page!" [47] The feedback is used for prediction. To recognize whether a certain coffee cup is yours [86] or to expect a certain feeling in your foot whenever you take a step [91], you are making a prediction. An invariant representation must be recalled from memory and compared to the most recent perceptions. [83] Hawkins says, "Prediction is not just one of the things you brain does. It is the primary function of the neocortex, and the foundation of intelligence." [89]

There is only one way to explain the speed of thought, using neurons wich require milliseconds to switch. The brain doesn't compute solutions to problems, but instead solves problems by memory operations. [68] According to the author: "The neocortex is not like a computer at all, parallel or otherwise." [69] Human memory uses "auto-associative recall" -- it retrieves stored patterns by means of their similarity to other patterns. [105] "We call this chain of memories 'thought.'" [75] "Thoughts are what occur when this model runs on its own; memory recall leads to predictions, which act like sensory inputs, which lead to new memory recall, and so on." [199]

The ability of the hierarchical system to run independently explains the common feeling that consciousness is some sort of "magical sauce" added to a "physical brain." [195] It accounts for the popular belief that a person has a soul that continues to exist after death. [199-200]

The new understanding of how the human neocortex works implies a number of useful applications, such as problem-solving hints [189-190] and means to deal with harmful social stereotypes. [204] However, the author dwells almost exclusively on one goal: that soon we will have intelligent machines.

What will these machines look like, and what will they do? While Hawkins hesitates to suggest a "vision" of the future, and avoids "the V word," [217] some general statements are safe to make. The author assures us that "unless we go out of our way to make them humanlike, they won't be." [216] Hawkins compares fear of intelligent machines to early fears of electricity, the steam engine, and the first computers. [213] The machine's model of the world will be very limited, e.g., to consider an intelligent transportation system: "The car needs to know about roads but not about elevators and airplanes." [221-222]

The intelligent machines will probably be given different senses than what humans have. Their senses may be sonar and radar, magnetic field detectors, weather sensors, and meters for society's use for electricity. [62, 227-230] Intelligence can use any senses because its task is "finding patterns in the world." [228]

An important feature of these systems will be "replicability." We will will need to train intelligent machines, and the first of each kind will have a long learning curve, after which we can replicate its software. [209, 226-227]

"Make it happen" is the author's advice to young people. "The Intels and Microsofts of a new industry built on hierarchical memories will be starting sometime within the next ten years." [235]

There is perhaps one point that is more important than the technology itself: In the epilogue, Hawkins repeats the observation of Carl Sagan that scientific understanding doesn't diminish the our sense of wonder about nature, but increases it. [234]

Book review by Mike Lepore for crimsonbird.com

Hardcover. 7-page index. Appendix proposing "testable predictions" to check the correctness of the author's hypotheses.[237-245] Bibliography, not in academic form, but in the form of a reading list with the author's comments. [247-252]