Review of "On Intelligence" by Jeff Hawkins
I have always been interested in Artificial Intelligence, how the brain works, and neural networks in computers, so I consider myself lucky that I stumbled across this book. Hawkins has a real passion, and it's fascinating to read how he has used that passion to get him where he's at and focus his success towards his goal of understanding intelligence. Back in 1980 while working for Intel, he tried to get them interested in funding research into AI; they surmised that computer technology wasn't really ready for applicable real-world AI (and they were right) and he had to look elsewhere. Then he applied to the Massachusetts Institute of Technology for an AI course, and was surprisingly rejected because he wanted to study biology of the brain as part of the course. I'm glad he stuck to his guns.
As for a bit of background on Jeff Hawkins: he founded Palm, and created the still popular Palm Pilot with its innovative Graffiti technology. This was after much research into hand-writing recognition, and also a little bit of human intuition in predicting how people would respond in having to learn how to write in a specific way in order to input data into it. He figured that if people are willing to learn an unintuitive thing such as touch typing, then there is no reason why they wouldn't want to learn something like Graffiti if it would benefit them. By making us meet the technology half-way, he also made a practical real-world example of how machines and us can feasibly relate. For example, when you talk to a friend you make certain assumptions based on regional accent, social and cultural background, current mood, motivations, and drives, etc. What this means is that it would be exceedingly hard for any artificially intelligent machine to come close to understanding us without knowing all of the socialogical and cultural background, as well as human drives and ambitions. This is obviously quite unnecessary for most applications, especially in the short-term, and there is no reason why we can't meet the technology half-way by doing what we do naturally: learn to effectively communicate. Personally I feel that if we are going to communicate to computers via talking, initially it would likely be something socially neutral like Lojban, but I digress.
Anyway, his approach to the book is to make it as simple and readable as possible. He doesn't bog the reader down with technical terms or theory, but rather skirts through the basics of what's required to get an understanding of his perspective. His style is quick to read, and although there are a few areas that inevitably end up heavy with technical detail, it never seems too much to take in or get your head around.
The main impetus of his theory is along the lines that our brains are dedicated to recognising patterns, and using those patterns to make predictions. For example, he talks about saccades (constant, minute repetetive movements of the eye that we're rarely conscious of) and how when we see a face, our eyes will automatically saccade from one eye, to the other, to the mouth, etc. When we look at a face, we don't actually see the whole face all at the same time, but it's made up of sequences of patterns that the brain assembles together into one coherent and stable pattern. We might see the same person's face upside down, or in different lighting or even making a face, but we'll still recognise it as that person. This is what he terms pattern invariance recognition, and is fundamental to the theory as he presents it. We can use recognition of simple patterns to make judgements about more complicated patterns, adding layers of complexity to the model as we go. A good illustration of this that he uses is how children learn to read by first recognising the letters. First looking at the letters, and learning the alphabet is time consuming, and uses most of your concentration, but eventually we move on to words, phrases and then sentences and no longer have to concentrate on the individual letters. This allows the higher areas of our brain to be freed up to allow more abstract patterns to be learned (patterns within patterns).
There is also a section on imagination and consciousness. I think it's partly to placate the more casual reader, as it isn't strictly fundamental to understanding intelligence - at least intelligence as he presents it - but it touches on interesting questions like: Are all animals intelligent? Why is human intelligence different? What makes us creative? What is faith?
He then goes on to investigate how intelligent machines might be created. In this section he argues that specialist chips will need to be created in order to enable intelligent machines, and that this is the single most difficult and expensive of the steps needed. This is contrary to a lot of prior opinion that it just required a sophisticated program and the right type of input, but when you think about it we have been working on AI with negligable success since computers were first invented, and realistically the layout of the brain is only superficially similar to a CPU of a current computer.
The last section looks at how intelligent machines might be used, and he makes a good case that they will be nothing at all like what popular sci-fi will have us believe. There was one point that I didn't quite agree on though, which was in an example of an intelligent car. Although his example wasn't intended to be realistic, he said that the car could be "trained" and the chip "fixed" for mass production. I'm not sure about this as it seems to go against how he previously presented intelligence: that the brain is constantly adjusting and adapting to its environment. My instinct is that this would prevent it from remaining "intelligent"; the car could cope with most situations thrown at it, but what would it do with any situation that's out of the ordinary?
Another question I found myself asking was: assuming that machines can be taught to complete certain objectives, how will they be motivated to complete those tasks in a way that best benefits our own interests? As soon as we start imposing certain restrictions and predefined procedures then isn't there a risk of ending up with something that could be coded better as a rigid application? The forum on the Numenta website has some similar questions and answers to these.
However, pedantism aside, this is one of the best books that I've read. It's short and easy to read (intentionally so), even if you know nothing about the brain, or computer science. It doesn't necessarily say anything new, but what it does do is put together all the disparate theories and principles into one coherent and tantalising whole. It's also made me think differently about a whole slew of things around perception, recognition, memory and learning.
Jeff Hawkins founded the non-profit company Redwood Neuroscience Institute, and co-founded Numenta Inc. where he is pursuing his dream. You never know, this could be the beginning of the most important thing to happen in the history of human development.
