Books about Brains and Machines

Being There: Putting Brain, Body, and the World Together Again, by Andy Clark

Of all the books I’ve ever read that related to artificial intelligence, robotics, cognitive science, neural networks, and the brain, I’m pretty sure this is my favorite. Usually, for me, these books are about filling in the blanks in terms of how other researchers see the world. I get frustrated because I feel that their assumptions are wrong, their view is too narrow, or they lack imagination. I need to know about their view, but it’s not one I share.

With Being There, on the other hand, I felt I’d finally found my intellectual home. I probably agreed with 90% of it, and felt that Clark and I were coming at the problem—how human intelligence works—from the same angle: that without considering the body and environment, studying the mind and brain won’t get you very far. At the time this was still a minority, emerging idea in opposition to the ‘top down’ approaches to intelligence that had dominated AI for decades. Clark (now at the University of Edinburgh) previously taught at COGS at the University of Sussex, a world center for this way of thinking about the problem, so he was ideally placed to tell the story: it’s now become mainstream.

Clark starts by giving us an introduction to ‘the new robotics’, describing what were then (1997) relatively new approaches to building intelligent machines. These were successful because they weren’t focussed on the idea of planning: where a robot would map the environment, detect obstacles, plan a route, and then start to move. This traditional AI approach was very slow and unreliable because by the time the machine had decided on it’s strategy and started to implement it, the world might well have changed. It would then have to start planning again. Instead, the approach pioneered by Rod Brooks (discussed in an earlier blog) showed the effect of concentrating on just letting the machine get on with it and seeing what happened: potholes, obstacles, etc. could be dealt with at the time. Can’t move forward: move backwards or sideways. Foot not supporting weight: put your weight on your other feet. In other words, use the environment to do your planning for you.

But don’t get the idea that this is a book about robotics. The bulk of the book is a description of how the behavior of most biological species is inextricably linked to their physicality and their environment. From babies learning to walk against the force of gravity (leg weight affects behavior), to slime molds coming together to follow temperature and chemical gradients towards better nutrients: he gives lots of great examples.

One of the ideas I found the most interesting when I read this the first time is how we organize our environment to make life easier for our minds. (The environment doesn’t just control us, we control it.) For instance, we put all the spices in the same place, which means we only have to remember the location of one thing, not several. We put a letter on the table on the way to the front door to remind us to post it. And if we get Alzheimers and lose our memories, we find we can still function well as long as we stay in our longstanding homes that act as external memories. He calls this scaffolding, and once the idea gets in your head you see it everywhere: in informational devices like books and computers, conveyor belts, ergonomic design. In the end, he even links this idea up with language.

The book is not really straight pop science: it’s a little more academic than that. But it’s a very good and fairly easy read nonetheless. I found it particularly striking because I had tried to read one of Clark’s earlier books with MIT Press: Associative Engines. There I found sentences the length of paragraphs and paragraphs that filled pages. I don’t know whether he found a great editor, took a writing class, or what, but there were no such problems here. For students and scholars, not only does he have citations (that can be ignored if that’s not your thing) but he also is really good about giving credit where credit's due within the main text. Reading this book is the equivalent of taking an engaging but easy starter course in this approach to cognitive science, neuroscience, philosophy, and machine intelligence. It’s rich and satisfying.

In a short review like this it’s not possible to do this book justice. All I can say is that those who find my blog interesting should go out of their way to read it.

On Intelligence by Jeff Hawkins

I have mixed feelings about what I consider to be 'celebrity' popular science books: being big in Silicon Valley and having something sensible to say about intelligent machines are two different things. In my view, however, Jeff Hawkins has paid his dues and deserves to be taken seriously. Though there is a lot wrong with On Intelligence, I believe Hawkins central theses—that much of what we call intelligence is based on the ability of our neocortex to make predictions, that the function of the neocortex is basic across sensing and actuation, and that this function can be understood in a relatively straightforward way—are all correct. If we're smart, some of us in the machine intelligence and neuroscience communities will take up his challenge to work on filling out this new theory.

However, I fear that this book was as much about promoting Hawkins new company as really pushing neuroscience forwards. If so, it's done a great job.

The book's biggest flaw from my perspective is that it puts up several unncessary barriers for those of us who are serious about the subject. The introductory chapters, for instance, are both too aggressive and too defensive. On the one hand Hawkins slights the last 30 years of research into artificial intelligence and neural networks in a far-too-general way, and on the other hand he talks up his own contribution in a way that made me uncomfortable (maybe I've been based in the UK too long?). Even if he's right, it's not polite (or necessary) to say it in the way he does. I don't know whether the writing style reflected Hawkins personality or not, but he didn't come across as a terribly generous person. This has nothing to do with his theories, of course, but much to do with the digestibility of the book.

I also found that the book was just a little too popular-science to be useful at times. I wanted more examples of real research, more cases where the theory tied into the reality of what other scientists had found. I also wanted more diagrams and fewer handwaving explanations. I felt particularly cheated because at various times we were promised that the sparse bibliography was supplemented by more stuff on the book website. This claim gave the book an undeserved credibility, as there are a total of four additional references on that site (at time of posting), hardly enough to have been worth missing out of the original book bibliography in the first place.

That said, the thesis is consistent with what I know about the brain (I'm no neuroscientist), well thought through, and well explained (if in a way that is not very helpful technically). Essentially, Hawkins argues that as information comes in from the senses it is sorted into increasing levels of complexity as it rises up the layers of neurons that make up the neocortex. So, for instance, lines become shapes then faces then specific people as you move up the hierarchy. At the same time, feedback from above filters the information coming in, so that the lower levels are specifically 'looking' for particular things. So if you see the top half of your friends head and then the rest is revealed, you'll be predisposed to see the bottom half you're expecting.

The next issue is that, if you don't (see what you're expecting), the information will be passed up the hierarchy until some level does expect it: either that, or it forces a new memory to form (Hawkins believes this happens in the hippocampus). Finally, he argues that this hierarchical traffic can go both ways: you can think about what you see (going up), or you can see (and do) what you're thinking about (going down). There's more to it, of course, but that's the gist.

Hawkins is right in that, as far as I know, that no-one has expressed this as an overarching theory in this way before. And I've little doubt that he's basically correct. However, all the pieces of this puzzle have been floating around for a while, and most researchers are only given the opportunity to work on pieces rather than the big picture. I felt a little more modesty wouldn't have hurt.

But perhaps that's the Silicon Valley talking. The book had two things in common with a patent application: a constant discussion of novelty with a distinct lack of details that would make it useful as a point of research for other people's work. A mixture of "insanely great" with "I'd tell you but then I'd have to kill you," if you will. I'm sure Numenta, the company formed to take advantage of the new technology, will do extremely well from this book.

The latter sections vary, and should be treated with care. After convincing us of his neocortex idea, he exposes us to several predjudices about embodied intelligent machines (robots), and  dismisses the problem of interconnections between neurons as hardly worth discussing. (It's a hard problem, and—many believe—a dealbreaker for artificial brains of any size). On the other hand he says some sensible things about consciousness and creativity

Anyone serious about machine intelligence, or brains in general, should read this book. Unfortunately, there is another, better, book out there to be written that really ties together Hawkins thesis with the evidence that supports it and explains his model well enough so that we could all go out and work on it... as he claims he wants to encourage us to do.