EPS Review #135 - On Intelligence

On Intelligence, by Jeff Hawkins with Sandra Blakeslee, Times Books 2004, 261 pp.

My friend Mike urged me to read this book, as you can see below. I am glad he did.

Hawkins thinks AI has failed to fulfil its promise, because computer scientists lacked interest in the brain. Turing equivalence made the brain seem irrelevant, and neural networks did not deliver enough. Hawkins feels the weight of Searle's Chinese Room argument (something I never did) and argues that a new brain-based computing paradigm will make real machine intelligence possible. He admits that, in the extreme, a brain could be modeled by a digital computer, but it is fair to say that speed is the issue there.

Speaking of speed, one of Hawkins' most striking arguments is the "one hundred step rule". You can recognise a cat in a picture in about half a second (while a modern computer can basically not do it at all). A synapse can fire and reset in 5 milliseconds. So whatever is happening when your brain recognises a cat, happens in only 100 steps. Sure the brain is computing in parallel, but the point is that something very different is happening from what computers currently do. And in some way, this problem is "simple" -- a shocking thing to say, but true in the context of the brain. Hawkins reckons that, when you catch a ball you are not solving differential equations, but are somehow simply remembering the solution, through an advanced form of associative memory.

Most of this magic is happening in the (neo)cortex, the folded outer part of your brain that is six layers deep. The cortex is relatively new in evolutionary terms, and only mammals have it. (This made me wonder about dinosaurs, and I found a web page in defence of bird brains. A dolphin has a big cortex, but only three layers deep.) Hawkins thinks the layers form a hierarchy, with sensory input at layer six, and time-invariant abstractions in layer one (with the hippocampus functionally at the very top, forming memories). There is a huge amount of interconnectedness among the layers, although it is mostly vertical in layers 2-6 and horizontal in layer 1. There is also lots of feedback downwards, sometimes ten times as much information feeds back as originally feeds forward from the senses. The author's theory is that much of this feedback is prediction of what is about to happen. (By contrast, neural nets have very limited feedback, and only at training time).

Hawkins thinks that prediction is the essence of intelligence. The Turing test defines intelligence in terms of behavior, but Hawkins points out that you can be intelligent just lying in the dark. Most animals control movement with the "old brain"; in humans the cortex has usurped this function, because in some sense predicting is doing. Likewise our eyes make constant saccades as the brain simultaneously observes and predicts the world we see.

The explosion of detail that Mike mentions is in the chapters on cortical columns (I should have mentioned that Hawkins maintains that cortex is basically the same stuff operating in the same ways whether seeing, hearing, feeling or whatever). I failed to pay very close attention to this part, because I thought that subsequent chapters would be on his experiments with duplicating this structure in silicon. I was mistaken. The last part of the book is some general predictions about a future with more intelligent machines, along with mechanistic explanations of consciousness, creativity and so on. There appears to be some research on his website (linked in first paragraph), but I have not read it yet. Maybe Mike will write a cortex simulator...

There are lots of other interesting brain facts: Einstein had more glial cells than usual, and more sulci in the parietal lobes. Women tend to have thicker cables connecting the left and right hemispheres.

I presume Sandra Blakeslee deserves the credit for the crisp and lucid writing. I think it will inspire many people to read more about the brain.

Mike wrote: Rituraj recommended a book called On Intelligence by Jeff Hawkins to me. I haven't read very much of it yet (I just got it) and I recommend it to you. He seems to have written up a number of theories about AI that match what I've been thinking. I'm very excited about the book (not something that happens often).

I'm in full agreement with the hypothesis by page 150, so I think it is worth it already. I should finish it shortly....

I finished On Intelligence, have you started it yet? Good book with some real problems -- first 80% is focused, suddenly too detailed, suddenly not detailed enough, suddenly philosophical (and not very thought out). The first 80% and sudden detail is the best. I still recommend it, since it covers much of what I've been thinking about.

Elise wrote: Sounds a little too scientific for me. But in a lighter vein, yet still attending to the brain, you might want to read, Stumbling on Happiness by Daniel Gilbert..........

Chris wrote: OK, that's one I'll read from your book list.

Sort of related thought:

I read recently that after the human genome was successfully mapped, researchers were surprised at how small the code was. They had predicted a much larger source. One reason could be that human code is not written in binary, but quaternary - if that's the right word - four possible amino combinations. That might account for it, but I wonder if there is more...

Another thought. I toyed with this at GS. Wall Street always needed a faster option calculator. As portfolios grew, the need was obvious. One solution I never got around to implementing was the sort of the OED Method of option calculators. I could calculate everything - once - and put it in a rather large 5 dimensional table. Meaning run a batch with every strike, vol, IR, etc combination and generate the premium. Thus any answer would be 5 blinding fast steps away. A good solution when you are speed limited but have unbounded memory - sort of like the brain - speed limited but unbounded(?) memory. Could that be the brain's trick, it has already calculated/"predicted" the cat. It ran as a batch the night before while we were all sleeping. And is that what sleep is, some down time while the batches are running?

And as a side note, does that explain deja vu? Your brain looks up cat images in memory to see if it knows this particular cat as input from the eyes. To do so, it compresses the data - maybe does something similar to a check sum and then does a look up to see if any of the stored images have the same check sum. A fast and but not necessarily unique way of looking something up. It is fallible. Two things can have the same check sum causing your brain to say "Aha!" even though it is just a false positive. You think you have seen it before but really you haven't. Deja vu.

And a really "out there" thought on the brain. If 10 years ago you had the ability to see 10 years forward, you would have looked at your current computer and thought gee, it must have a very large hard drive because look at all the info in there. Every book every written, every news story, every song, even TV broadcasts.

But of course, from today's vantage point, we can see that your hard drive hasn't gotten larger at all - you are just connected to something beyond the boundaries of your computer's hard plastic shell - the internet.

Back to the brain, could it be that the OS in your brain isn't a fully loaded version of Windows - but instead a pared down version of Internet Explorer? Meaning to recognize that cat, you are surfing and getting the scoop from elsewhere? I'm not implying God in my old age, but the fact remains that the source code for Human ver 1.0 is smaller than it should be....

And that would mean that sleep is actually people uploading back to the mothership... (kidding).

I think it would be the ultimate interesting programming project to write the brain's OS. Boy that would touch upon everything. I suspect you'd have to program in pain and pleasure to motivate the program to learn. Heck, you'd probably have to go all the way back to some newly invented first principle and understand why all these carbon and hydrogen atoms want to cooperate with each other in the first place - I still don't get that.

And is it true that the human body gets 21 grams lighter at the exact moment of death? Hmmm, does your computer get infinitesimally lighter when you reboot?

I wonder if true gut level assembly programmers have ever taken a look at DNA and tried to figure out the op codes and the simple code blocks like loops and conditional statements. There should be a Guggenheim Grant awarded for this...

Mike wrote: A couple of observations - The cells in the human body are polymorphic - proteins are expressed and inhibited based upon external forces. The code maybe small but the complexity is still enormous. From simple rules come complex systems.

In terms of fast option calculators, the system FUT which we wrote for the floor of the pit in Chicago when I was in equities (circa 1989) did exactly what Chris is describing. Pre-computed daily (or when events got hairy) five dimensional matrix, accurate but fast interpolation. The system worked great and still sits in my directory at GS and works (if the directory still exists).

I think déjà vu comes from short-term predictions engendering a feeling and then coming true. But Chris's idea sounds very interesting to me and may make more sense.

I have the following questions raised by the book --

How is time serialized? How is something temporal stored in the spatial areas of your brain? Are patterns recognizable by a single neuron? Is it more complex than an and gate (with special temporal stuff)?

How does pattern matching actually move from higher regions down to lower regions? I think Hawkins is wrong on the mechanism and that what happens is a whole area of the brain works at a gross level at first and then is constantly being refined into smaller regions. This would mean that data or patterns don't have to be moved from one place to another in the brain. There are no obvious mechanisms described for moving whole sections of the brain around.

Another phenomenon not even touched upon is if/how/why new axons and dendrites grow. Why are there different neurotransmitters? I personally would have preferred discussion in these areas over the philosophical musings at the end of the book.

I had a different approach in mind than a cell level simulator. The book may have changed my design quite a bit, I'm still deciding. I was going to abstract patterns into higher level objects instead of deriving patterns from continuous stimulation of billions of cells. I think the brain largely does what I want to do, but the cellular mechanism definitely approximates reality better.

So, I'm not sure what I'm going to do.

Jeff wrote: I took several neural sciences courses at Brown -- fabulous stuff indeed.

I replied to Chris: Hawkins makes the point that for most animals, like a bacterium heading in the direction of an increased nutrient gradient, intelligence is hardwired in the DNA. But for humans, the DNA codes for more cortex, and then that organ, with the aid of culture and experience, produces a whole greater magnitude of intelligence.

I like your mystical musings and figure they would be well served in a SF story (if they aren't already). Sadly, the mechanistic explanations seem more likely. Did you ever read The Emperor's New Mind where Penrose supposedly says the brain uses some quantum magic and so is fundamentally something more than a Turing machine? I have the book but have not read it.