Some Thoughts on On Intelligence by Jeff Hawkins
by David Ashley
April 3, 2007

I loved the prologue and first few chapters of this book. Hawkins is passionate about developing a theory of mind, about understanding how the brain works. He wanted to pursue this but met with resistance -- no one wanted to finance this research, in the 1980's it was too risky. Hawkins writes a lot about how wrong the current approach to AI is. I agreed with everything he wrote. It was very refreshing reading a lot of the same thoughts I've had myself.

Then the book turns south. Hawkins has developed his own model for how the brain works. He's the first to admit it's probably wrong and in need of work. The basic theme of his model is that the brain is always trying to predict what's coming next.

There is a lot of detail about his hierarchical model, how lower levels pass their analysis of low level patterns up to higher levels, how higher levels look for higher order patterns and pass their analysis up to still higher levels. Then at the top there is the reverse, commands get sent down the levels to initiate simultaneous playback of patterns out. When the predicted patterns match the incoming ones, at every stage, the system doesn't call attention to itself. Only when there is a disagreement at some level is a higher level notified of the problem. Supposedly the higher level can adapt and get things back on track.

I'm paraphrasing. Hawkins goes into a lot more detail. I only read it with enough attention to get the gist of the theory. Basically it's nothing like my own theories, so I only examined it to look for new insights and also for the purpose of poking holes in it. I applaud Hawkins for his intricacy and detail, how he's written it all down and published a book about it for everyone to gawk at and criticize. However in the end I simply can't get behind his approach.

Hawkins doesn't go into any detail about how the system learns these patterns. Instead it's sort of given that the system starts out already having the patterns programmed in, and once that has been done, it all works like... this. Hawkins' hope is that by putting something down it can act as a starting point. He encourages young scientists to pursue brain theory as a career, to take chances, start companies. He freely gives away his ideas in the hopes they'll assist entrepreneurs in forming businesses.

One thing I found particularly ironic about the schism between the first part of the book (which complains about the gridlock present in current GOFAI (Good Old Fashioned AI such as symbolic logic and neural nets)) and the second part where he presents his theory, is that he himself does exactly what he complains about early on.

Specifically, in chapter 2 on neural networks Hawkins describes an analogy. The idea is to imagine people reverse engineering a computer, finding it's built up of transistors, then discovering that putting a few transistors together in a certain way makes an amplifier. Then this spawns new industries that then make gadgets (and money) such as radios and tv's. However the problem was that the original intent was to be able to reproduce the computer -- and that wasn't done. The point is that people looked at neurons in the brain and went off and invented neural nets -- which are useful (supposedly) for some purposes, but are nothing like actual neurons in a nervous system.

He makes a wonderful point -- one I agree with wholeheartedly. But then Hawkins himself did exactly what he warned us against. He developed an incomplete theory of brain function, then went off and started a company to exploit the theory. The company is Numenta. He also started an institute intended to develop a complete theory of the brain, called the Redwood Neuroscience Institute. Now, I think he started the institute before Numenta. The institute seems to have been absorbed into the University of California at Berkeley and Hawkins doesn't seem to be directly connected to it as a driving force anymore. This separation seemed to coincide closely with the founding of Numenta. I'm not sure of my dates + "facts". It's just my intuition from piecing together the bits and pieces I've absorbed in the limited time I've invested.

So Hawkins complains about people "selling out" and then goes and does it himself! Note Hawkins is probably rich, rich, rich from success at Palm Computing. He also probably made a few bucks from "On Intelligence" the book, after all I acquired a copy somehow. So he doesn't need money.

Based on the little I know about the story, I consider Hawkins to have committed a sin against his original passion. He's already rich. He's a smart fellow. The correct thing would have been to stick around at this Redwood Neurosciences Institute he founded, direct its efforts, and perhaps even spend a bit of his own time developing his theory further.

Instead, Hawkins seems to have changed his focus to making a buck by forming Numenta and trying to devise ways that his model of intelligence can be implemented to solve real-world problems. This is a trajedy, for two reasons. Firstly his passion would have lent a necessary driving force to the Institute. I suspect in passing it off to UCB it will have lost whatever sense of urgency it started with and will become just another think-tank, producing nothing useful, and focusing mainly on locating grant money (ala The Santa Fe Institute). Secondly, Hawkins' theory is wrong. Trying to go for the $$$ with Numenta means rather than form a complete correct theory, Hawkins will instead focus on trying to squeeze out whatever money can be had with yet another incorrect model of intelligence.

Numenta will probably make money. We'll probably see some very interesting products come out of the place. But real deal intelligent machines? Forget it.

Now I'll provide a bit of detail on why I think Hawkins' theory is wrong.

The main thing is, it's incomplete. Where does learning fit in? How does the hierarchy get itself built? It is just a mechanism to record dynamic patterns and play them back. How does this turn into an intelligent machine? Hawkins presumably leaves this as an exercise for the reader. But the critical essence of intelligence is learning itself. Hawkins has no theory. It's just a description of one of the many symptoms of intelligence -- pattern prediction -- and enough details of an algorithm to implement it with typical digital hardware of the day.

When Hawkins speaks of symbols getting passed up and down the hierarchy, when he speaks of sequential patterns (as in A, then B, then C, then D occuring in rapid succession would become an ABCD pattern), right away I know he's blown it. The brain doesn't work that way. The fundamental building blocks in the brain are not symbols. Trying to define symbols and work with them is a throwback to GOFAI. It's a dead end.

The learning comes first. The brain learns to create and manipulate symbols. The symbols are a result of intelligence -- not the basis. It's like this: I have a toolbox and I can fix cars when they break down. A savage straight out of the jungle might see me fixing a car, then might decide he wants to do that. So he steals my toolbox, thinking the toolbox and the tools inside are the magic I have that let me fix cars. But it won't work. The tools are not the magic. Symbols are tools.

I don't have a complete theory of the brain either. But at least I know enough to keep working on the problem until it is complete. One thing I'm sure of is that intelligence is an emergent phenomenon. As such I see GOFAI and Hawkins' approach as essentially identical -- attacking symptoms of intelligence instead of hitting the root causes.

Having said all the above, I still very highly recommend Hawkins' book On Intelligence. If you're young and thinking of going into a career in AI, please read this book, at least the first few chapters. Read the first part because you really want to know what you're getting yourself into. Hawkins blasts the traditional approaches to AI quite effectively. Don't go down the same path academia has -- it's a dead end. Next, if you do wish to go on and read the whole book and get introduced to Hawkin's theory, take it with a healthy amount of skepticism. Just because Hawkins is smart and rich and successful doesn't mean his theory is the truth.

Access count (e^i)