Customer Reviews

On Intelligence

5 star:		(86)
4 star:		(31)
3 star:		(9)
2 star:		(4)
1 star:		(5)

 

 

It is not very often that you encounter a book that alters, not simply what you think, but how you look at the world. On Intelligence is such a book. Jeff Hawkins develops a perspective on intelligence that makes sense of much of what I have discovered about learning over the past twenty years. His focus is on a unified model of how the cortex works, but in truth you do not need to have deep interest in neurobiology to see the power of the model. The book is very clear and readable, something I have learned to associate with Sandra Blakeslee's deft touch (see, for example, Phantoms In the Brain, by Ramachandran and Blakeslee). The heavy lifting occurs in the lengthy sixth chapter, "How the Cortex Works." You might want to skim this chapter or even omit it entirely on your first reading. It is well written, but requires a very thoughtful reading. The model Hawkins develops in this chapter underpins his view of intelligence, but it is not necessary to grasp the details to appreciate the power of the vision. If you have the slightest interest in the role of the brain in making us who we are, you owe it to yourself to read this book. I couldn't recommend it more highly.

The accolades previous reviewers have lavished upon this book are all fully deserved. It is not, however, "the first time all these bits and pieces have been put into a coherent framework". The work of Stephen Grossberg explored all of these themes in the 1970s. Unfortunately Grossberg expressed his key insights in systems of differential difference equations that few could understand and fewer still could build upon or contribute to.
To his credit, Hawkins does cite Grossberg approvingly at several junctures in his argument, but he fails to take into account several of Grossberg's greatest insights into neocortical processing: his theory of how serial processing can be accomplised in a parallel anatomy and his theory of "rebounds". The latter is especially important since it explains how new memories are prevented from overwriting old memories. For example, when I learn a second language, it doesn't overwrite my first.

These criticisms, however, are in no way meant to detract in the slightest from Hawkins' superb book. It is an eminently readable account of neocortical computing, and correct in all its broad brush strokes. If you are as beguiled by "On Intelligence" as the other reviewers in this thread, my purpose is only to alert you to the even deeper wonders that are to be found in Grossberg's work. As I have said, his work is difficult, but his 1980 and 1982 Psychological Review articles will provide good entry-points. Those of you with an interest in brain and language will find an even better second course in neocortical computing in Loritz' "How the Brain Evolved Language" (Oxford University Press, 1999).

Jeff Hawkins is the man who was the architect of the PalmPilot, the Treo, and invented Graffiti, an alphabet for inputing data to a computer with a stylus. But this book is about his other love, the deciphering of the code that makes the human brain work. There is nothing like a big, important puzzle to get the blood working, and mine was powerfully pulled along . With the human genome project's sequencing of human DNA nearly completed, understanding the brain has got to be the most important scientific undertaking one can think of. Hawkins easily persuades us that there is a burning need for a "top down" model for the brain that can play a role something analogous to the Central Dogma of molecular biology, which guides and organizes research, prioritizing the myriad of possible tasks into something like that required for the logistics of a conquering army's march through an alien land.

He also persuaded me that he has some important insights of that model that I found tantalizing, new and exciting. His central model concerns the role of the cortex in producing intelligence. He makes the case for a central dogma he calls "the memory-prediction framework." This idea says that the cortex is a machine for making predictions for temporal sensory patterns based on memories of past patterns. The prediction algorithm carried out in the cortex is the same for all of the senses of vision, touch, hearing, etc., which accounts for, among other things, the basic physiological uniformity of the cortex, and the plasticity of the brain in adapting to such problems as blindness or deafness.

He argues that since the "clock" of the brain operates at a tick-rate on the order of 5 milli-seconds, and most of the functions of the brain (e. g. recognizing that a picture of a cat shows a cat) are carried out in less than 100 ticks. From the time that light enters the eye, to the time it takes to signify recognition takes less than a second. A computer would take billions of instruction steps, and even the fastest parallel computer available would not do it in less than millions of steps. So the brain doesn't really "compute" the answer, it retrieves it from memory, which requires far fewer steps than the computation. Sounds good to me.

His explication of the memory-prediction framework is clear and accessible even to the uninitiated like me, though I found some of it in the middle pretty heavy going. But this is something like reading Watson and Crick's paper on the structure of DNA. The part about turning the diffraction diagram and other insights into a workable model was a little above my head, but I could still see the importance of the answer, and how it addressed the problem of replication and how it gave clues as to how to "read the genes." I can only grasp part of what Hawkins has done, and I can see that there is still a long way to go. But I can still jump up and down about it!

The early parts of the book (up to around p 60) were a great read and convinced me to buy the book. But when Hawkins finally laid out his "big ideas", I was deeply disappointed. Hawkins spends considerable space claiming that AI researchers hack up algorithms based on the "how do I do it" approach. He suggests that "real" intelligence requires memory-based hierarchical models.

What is especially frustrating to this AI (specifically vision) researcher, is that Hawkins does not seem to be aware of any AI research that has been going on in the last 15 years, during which data-driven learning approaches have become standard. I was merely suspicious of his ignorance until I checked his bibliography, in which the most recent technical AI citation was from before 1990.

Furthermore, Hawkin's theories on the brain are largely unsubstantiated. He states that his ideas were largely sparked by one dated paper that other researchers have largely ignored - probably for good reason. For instance, he claims that, since different parts of the brain have a similar physical structure, they must function similarly. This is very oversimplistic.

Nevertheless, I did find parts of the book to be entertaining and appreciated his view on the brain's role as a predictor. Although I do not think that I completely wasted my time in reading this book, my time could have been better spent reading something else. Therefore, I recommend this book to non-scientists who want to read about the brain but aren't particularly concerned about the accuracy/usefulness of what they read. Just be a very critical reader and be careful not to be smacked in the course of all the hand-waving!

There was plenty of buzz last week about the new company - Numenta - that Jeff Hawkins (inventor of Graffiti and the PalmPilot, Visor, and Treo products) and Donna Dubinsky (CEO of Palm and Handspring) have started. It was coincidental that I was reading Hawkins book - On Intelligence - which describes his theory of intelligence, the working of the brain, and how he thinks it will lead to the creation of truly intelligence machines.

I haven't spent any time studying neural science, the brain (my biggest effort was probably not very successfully grinding through the Scientific American issue on Better Brains), or any of the contemporaneous efforts at "next generation Artificial Intelligence" (I was at MIT in the 1980's during the peak of the last wave of AI research and subsequent commercialization attempts - I fondly remember being amazed at Symbolics - they are still around in a new incarnation called Symbolics Technology - Macsyma has been hard to kill off) .

So - I don't know much about brain research, theories of intelligence, the biology behind it, or much of anything else. As a result, I thought On Intelligence was superb. I don't expect that it's right (nor does Hawkins) - he's clear that it's a framework and work in process (as it should be). I found it extremely accessible, very provocative, and mostly internally consistent (which is important whenever you are trying to learn about something you know very little about - it can be wrong, but at least it hangs together in a way you can understand it.)

The book and theory is based on the work being done at the Redwood Neuroscience Institute, of which Hawkins is the founder and director. Beyond just doing research, part of RNI's mission is to "encourage people to enter and pursue this field of research." Hawkins is consistent in his message in the epilogue of his book where he says "I am suggesting we now have a new more promising path to follow. If you are in high school or college and this book makes you want to work on this technology, to build the first truly intelligent machines, to help start an industry, I encourage you to do so. Make it happen. One of the tricks of entrepreneurial success is that you must jump head first into a new field before it is one hundred percent clear you can be successful. Timing is important. If you jump too early, you struggle. If you wait until the uncertainty lifts, it's too late. I strongly believe that now is the time to start designing and building cortical-like memory systems. This field will be immensely important both scientifically and commercially. The Intels and Microsofts of a new industry built on hierarchical memories will be started sometime within the next ten years. It is challenging doing new things, but it is always worth trying. I hope you will join me, along with others who take up the challenge, to create one of the greatest technologies the world has ever seen."

Hawkins thoughts and writing are fused with his obvious entrepreneurial energy. He approaches things as an ultimate pragmatist (unlike so many scientists, his examples and analogies are extremely understandable - very reminicient of Richard Feynman), an outsider (he acknowledges that mainstream brain research has huge problems with many of the things he is saying), and recognizes that any fundamental breakthrough typically requires a paradigm shift in thinking about the specific domain.

If you are an entrepreneur who likes to challenge yourself intellectually with things you know nothing about, you'll love this book. If you are a brain researcher or scientist, you'll probably be frustrated, but it'll stretch you in good ways. If you are a brain expert, you'll probably hate it. In any case, it'll be fun to watch what Hawkins, Dubinsky, Numenta, and RMI do next - remember, they're the ones that brought you the Palm Pilot / Handspring Treo based on the revolutionary notion that humans should learn to write different (e.g. Graffiti), not the ones that brought you the Go Whatever or the Apple Newton who thought that the computer should be able to recognize your handwriting.

RECOMMENDATION:
===============
This is one of the few books to posit a theory of human and general intelligence, and the only book of the few that is clear and well written. I think it is seminal will re-ignite interest and activity in building intelligent machines. A pleasant interesting read even for those not working in the field.

COMMENT ON CONSCIOUSNESS:
==========================
In one sense, this is a side issue since one can probably build intelligent cars, vacuum cleaners and search agents without consciousness, but in another sense it's a crucial aspect of our experience. Hawkins claims that consciousness is just what it "feels" like to have a cortex. I differ. My guess builds precisely on Hawkins suggestion that the cortex is a generative (my word, his is associative completion) hierarchy. That is, we synthesize/simulate the external world inside our head. But, we're a social creature and place a lot of value, evolutionary and otherwise on being able to imagine/simulate the mental state of other people ("my boss will be angry if I do that", "she likes me", ...). Yet, as a mater of simple functioning, we must also simulate ourselves in the world to know how to act. In my mental world, I simulate myself when I consider whether I can squeeze through a gate or lift a weight. When our simulation of mental state became grafted to our simulation of self, I think consciousness came about as an epiphenomena - consciousness is our simulation of our selves, of our own internal state.

ONCE YOU GET THROUGH CHAPTER 6:
===============================
Some holes which might exist either in my brain or in Jeff Hawkin's theory:


ATTENTION:
P-173 Attention gets pretty short shrift in the presented theory down to an alternative, hierarchy bypassing pathway in the Thalamus that gets turned on by higher regions if unexpected events occur bellow or the higher region is directed externally - the last is somewhat circular reasoning: attention is turned on if attention is directed. John Reynolds at Salk has been studying visual attention in monkeys and is finding evidence of boosting or diminution of contrast is what visual attention is doing so that visual items win or loose the inhibitory competition between features and that this is perhaps what lets some items rise up to conscious notice.

Attention is fairly sequential and substantially bottlenecked for what it can process (see "change blindness" illusions http://viscog.beckman.uiuc.edu/djs_lab/demos.html ). In many of these illusions, you don't notice when huge portions of the visual scene change, items appear or disappear etc.

CEASLESS RECODING OF MEMORY:
Hebian learning is great, except that it also unlearns equally well. I quote Grossberg's term for the problem in caps above. Memory needs some kind of gating mechanism or it will rapidly turn into mud. Either memory is unidirectional (connections start out high and only shrink, or starts out low and only grows), and/or there is a gating mechanism that isn't well explained here. What stabilizes learning? P-136: a purple "bucket" became "indigo" (or a page earlier, orange is placed in "red"). First of all, this can shoot down a whole painstakingly learned hierarchy of learning above - in general, a bad move. Ever done visual tracking algorithms? - if you allow your template to adapt a little bit in say tracking a face, pretty soon a little bit of background "wall" starts entering the template and pretty soon "wall" becomes your (very stable) "face" template. The same thing will happen here - color buckets will randomly turn into each other, drift around - chaos. Just like our legal system, most new rulings should have very local effects and only very rarely will something ripple changes through the larger system. If this happens too often, the whole structure collapses.

Finally, this ignores all the critical period stuff in learning. Some things are laid down early and in clear order and they don't seem to change and if not learned early just cannot be learned. Famous study of this is "Kitten in the carousal" where kittens are raised in the dark and only get to walk in the light in short intervals where one cat can move but is mechanically yoked to another cat who sees the exact same things, but is stuck in a carousal (a little box) so that it's leg movements don't control it's movements. If this is done too long, the poor kitten in the carousal never learns to see (depth) at all! Even when let go into the light. If let go early enough, it will learn to see normally.

Long winded, but: Seems to me that some basic categories and features must be developed early and not allowed to change in order to have any chance of building a larger structure over them.

TIME:
Where did it go? I see sequences, but not timing - you can't control your muscles without actual timing, not just sequence. In fact, time itself is yet another unstated sense. There are clearly integration rates that are learned and used in recognition, planning decisions etc.

INVARIANCE/FEATURE SELECTION:
I still don't get exactly how invariance is found by this architecture beyond things that can be predicted which is somewhat of a tautology - yes, invariant features make your life easy, but beyond dumb luck, how do you find them? When you identify a dirt road by parallel tracks in the soil, what inside you is discovering the cross ratio projective invariance? How did we learn brightness normalization? Color constancy? Some of this stuff involves tricks in active diffusion of color information from edges and clever local integration. Is this learned or built in? Insects must have to deal with this and must be born with it. How do they do it?

GATING:
P-158: Thinking of doing becomes doing. Yes, but how to you stop this from happening? Indeed, how do you start one invariant representation of say the Gettysburg Address from being spoken, written by all limbs and done in interpretive dance once you think to do it?

Minor nits:
P-71 While I believe that the fundamental unit of processing is a kind of sequential associative memory, the fact that you think or recall serially doesn't prove this - perhaps you can recall everything in you house at once, but internal or external output is one thing at a time and nearby things just have a scotch more support.. Detailed motor execution is more compelling.

I could have done with a final summary 2 side to side page cortical sheet diagram with Thalamus, Hippocampus, and at least 2 layers of hierarchy with all the basic communication channels and their direction shown, even better with text referencing where these things were described.


Gary

I am a neurosurgeon and I picked up this book with a great deal of scepticism because in the past, neither through my professional studies nor through reading many popular books, have I really been able to answer some fundamental questions regarding our brain. Questions like how do we think? what is imagination? what happens during the so called "a-ha" phenomenon, when you suddenly understand something that you did not a moment ago and many other such questions have been plaguing me for years. And believe me, there arent many satisfactory answers floating around.

Hawkins has made a fantastic contribution by giving us A model to think about these questions. His memory-prediction paradigm is very attractive intuitively because it automatically explaines so many facts about our brains and their evolution that other theories just ignored. But even when tested in hard scientific experiments, I predict that the basic structure of his arguement will remain intact, though details may differ.

We still do not know a lot about the architecture of the brain, the way neurons are connected to each other and the way brains develop their enormous computive capabilities. As we learn more, it is likely that Hawkins paradigm will be refined, but in the long run, we will owe Hawkins gratitude for allowing our brains to understand how our brains work!!

Jeff Hawkins is an entrepreneur and computer expert, responsible for the invention of the popular device known as the PalmPilot, as well as the Treo smart phone and other gadgets. He is also interested in the human brain and how it functions. So it should be no surprise that he has chosen to bring together his two main interests -- computers and the human brain -- in a book entitled "On Intelligence" which presents a new theory about how the brain works and how we can finally build "intelligent" machines.

Of course, discussions about computers, artificial intelligence (AI), and the possibility of building intelligent machines have been plentiful for many decades. The English mathematician Alan Turing, an early developer and innovator in the field of digital computers, best known for the Turing machine and the Turing test (both concerned with the relation between computation and mind), proposed a criterion in 1950 which would determine whether or not a machine can "think." A machine can think, he said, if its replies to questions are indistinguishable from those of a human being. With the declaration that "the human brain is just another computer," the field of artificial intelligence was launched.

Turing's declaration, however, became controversial and was criticized by both scientists and philosophers, especially those working in the areas of learning psychology and philosophy of mind. Turing's position, now known as "strong" AI, was especially criticized by John Searle, a philosopher and cognitive scientist who created a thought experiment, called the "Chinese Room" argument, which demonstrated that, while the computing device could indeed reply to questions in such a way that made it indistinguishable from a human being, it had no "understanding" regarding its replies, no "meaning" was attached to its replies, and it was not really behaving in the same way that a human being does. Turing's test was shown to be faulty and misleading.

In this book, Hawkins goes beyond Turing's ideas and Searle's discussion of the matter, and argues that intelligent machines can and probably will be built, but that a basic understanding of how the brain actually operates is fundamental to the development of such machines. The brain is not a computer, the author claims, but a memory system which makes predictions based on memories resulting from the interaction of events and their relationships. "Intelligence" is defined by Hawkins as "the capacity of the brain to predict the future by analogy to the past." And the first necessity on the way to building an intelligent machine is to understand how the human brain actually works, a subject to which he devotes most of his book. The reader will learn a lot about the evolution of the animate brain, including a lengthy discussion of neural networks and how the neocortex works. The author provides credible information and a compelling framework with which to understand brain activity.

Be that as it may, "On Intelligence" is not, despite its arresting title, a treatise on "human" intelligence. First, and I am not one who usually quibbles over definitions, his definition of "intelligence" is too limiting and his book should really be titled "On Animal Intelligence" or "On Machine Intelligence" or, maybe better, "On Computer Intelligence." I would argue that when it comes to "human" intelligence there is a lot more involved than merely "the capacity of the brain to predict the future by analogy to the past." In a "strict" sense of the term, human intelligence is an activity of the "intellect," that cognitive faculty of the mind as it operates at higher abstract and conceptual levels, and thus refers to universal ideas, judgments, and reasoning. These "intellectual" activities, which we philosophers in the classical realistic tradition call "intellection," are virtually ignored by Hawkins. Yet these are the essential activities which make us members of the class of human beings.

Second, Hawkins concludes his discussion of consciousness and creativity (Chapter 7) with an interesting paragraph. He states:

"By now, I hope I have convinced you that mind is just a label of what the brain does. It isn't a separate thing that manipulates or coexists with the cells in the brain. Neurons are just cells. There is no mystical force that makes individual nerve cells or collections of nerve cells behave in ways that differ from what they would normally do."

No, I am sorry he has not convinced me that "mind" is merely a "label" for what the brain does. Actually, he never defines the term "mind," so it's hard to know what he is really saying. The traditional definition of "mind" as "the conscious knowing subject or the conscious knowing part of the subject" seems to me to be pretty clear and has nothing to do with a "mystical force." It seems obvious to me that "I" am not my "brain." My brain is a physical organ which permits me to have an "I" (ego) in the first place, but I would argue that my "I" is not a label for what my brain does.

Third, if I am to infer that he equates "mind" and/or "intellect" with "brain," then his basic thesis regarding human intelligence rests on plain old-fashioned metaphysical materialism and, probably, old-school psychological behaviorism. I would argue that both these philosophical positions have pretty much been discounted today, as these "theories" have been unable to explain and account for the vast array of human activities, both objective and subjective, which all members of the human species experience in ordinary life.

Nevertheless, even with its shortcomings, I found the book an interesting read and would recommend it to all those interested in the subject of "intelligent" machines and the future of the digital computer. I just want to warn those readers who may take Hawkins uncritically that there are some philosophical implications here that are important and which the author does not directly address. It is well-written and most readers with any "human" intelligence should find it an easy-to-understand discussion of a relevant topic.

It is pretty clear that there are two classes of reviews for this book. One class, typically written by lay people, believes it to be the best research available on how the human brain truly works. Scientists, however, view the book a bit differently.

I am a researcher in robotics and specialize in developing control systems for autonomous robots. My company builds robots that can move around, and that have arms with which to pick up objects, all working without human control. Vision and touch are the senses used by our machines, combined with biologically inspired computer algorithms, to get the job done. Most of my work, like that of Mr. Hawkins, focuses on thinking about how animal brains might work and applying those thoughts to real systems.

I believe that Mr. Hawkins is a very sharp guy, and he describes his ideas about how the brain works with great clarity. He is outstanding at creating buzz. But, with all due respect, I believe that he doesn't even know what he doesn't know when it comes to building systems that work in the real world. The book reads as if the theories espoused are based on science, but they are really based on the author's conjecture. True, it is reasonable conjecture, but not fact. Software reportedly has been written based on these theories that is capable of recognizing hand drawn objects. I have not found any papers to review concerning this technology, but similar technology (e.g. OCR) is already available that is robust when recognizing hand drawn characters so this is not yet a tremendous breakthrough. Basically, working with 2D images is relatively easy, working with a computer generated 3D world is 10x harder, working with real imagery in a constrained environment (in a lab with controlled lighting, etc.) is 10x harder still, and working outdoors in the real world is about 100x harder than that. Current technology for autonomous robotic control and object recognition is not based on techniques of classical AI, but is in fact based on pattern recognition/matching techniques essentially similar to what Mr. Hawkins proposes, including the idea of prediction.

On the one hand, I applaud the author if this book inspires other people to enter the field. On the other hand, readers are cautioned that this is a "popular science" book and does not represent any great breakthrough.

This is an interesting book, but I'm not at all convinced of most of its major theses. There are way too many statements like "I believe xyz" in the book, and way too few along the lines of "Empirical evidence shows xyz." Hawkins seems to have committed himself to certain dogmas, many of which are probably oversimplifications. For instance, he insists that all the areas of the neocortex are essentially instances of the same software; for a completely contrary view, see Steven Pinker's The Language Instinct and How the Mind Works. Pinker, unlike Hawkins, starts by painstakingly laying out the evidence for the things he really knows empirically are true, and only then indulges in wild speculation.