I’ve been meaning to read this for a while, and I added it to my last Amazon order, but didn’t get around to reading it until a few weeks ago. Jeff Hawkins was one of the driving forces behind Palm and Handspring, and now that he’s set for life, he’s indulging his childhood dreams of trying to understand the brain by starting the Redwood Neuroscience Institute, which is apparently now the Redwood Center for Theoretical Neuroscience. In this book, he pulls together a layman’s overview of neuroscience literature that he finds interesting, and then espouses his own theory of how the brain (or at least the neocortex) works.
Here’s the basic idea of his theory. The neocortex is composed of pattern-recognition elements that are wired to remember events that occur together. It’s a hierarchy of pattern-recognition elements that breaks people’s perception of their environment into manageable chunks. In other words, when I look around the room, I don’t see ten million pixels; I see my desk, the computer, the wall, etc. Even if I look at my desk at different angles, my brain perceives it as a single object.
Another aspect of this is that these elements are learning new patterns all the time. When we learn to drive and first get out into traffic, it’s terrifying because our brains are overloaded trying to filter all of the myriad visual information around us. As we grow more used to the speed of traffic and learn what’s relevant, the visual load is automated and pushed down to a subconscious level of the hierarchy. The same holds true for recognizing positions on a gameboard. None of this is particularly novel (I espoused a similar idea in my cognitive subroutines proposal).
The novel bit is that Hawkins noticed that our brains do more than perceive – they are actually continually making predictions. Here’s an obvious example: when you’re scanning your home, you can notice when something is NOT there. How is that possible? It’s not there, so you can’t see it. But your brain has developed a model of what IS there, and is making a prediction for what it should see, and when something doesn’t match its prediction, it alerts the conscious mind that something is wrong. This makes a ton of sense. Our brain adapts to the familiar, but if something changes, it needs to turn all of its attention to understanding why there’s a discrepancy. I thought this insight alone made the whole book.
As an aside, this also explains why most people suck at estimating probabilities. Our brains are wired to remember the abnormal and outlandish because they break the routine patterns that we have learned. We don’t remember the 99% of the time when things go as we expect them to, because it’s all handled subconsciously. So we significantly overestimate outlandish risks because they break the pattern and come to our conscious attention.
Another discussion that I liked was Hawkins’s description of “invariant representations” (which I allude to in my post on localized generalities). Basically, because the neocortex is hierarchical and each level is always making predictions, each level can notify the level below it what it should be looking for. In other words, if one level keeps track of things in my office, it can notify the level below it that it should be looking for a desk, and that it should figure out how to interpret the raw sensory input in such a way that it looks like a desk.
As another aside, this also explains why we often see what we expect to see. Our entire sensory system is designed around the principle that it should adapt its interpretation of raw sensory data to match what the levels above it think it should be seeing. This applies not only to physical things like desks, but also when we see patterns in random data, or “interpret” data in such a way as to support our point of view. Our brains are wired that way.
I thought the book was decent. The predictive aspect of the brain and the discussion of localized generalities were “Oh, wow” moments, as I immediately saw how they filled in gaps in some of my theories. Most of the rest of the book was an explanation where he handwaves how the current understanding of the neocortex can support his theory. There’s some minorly interesting stuff in there about how the various neocortical layers are connected in a way that might be hierarchical in the way he suggests, but that’s mostly of relevance to the neuroscience geeks.
I’m mostly kicking myself after reading it, though. I was moving along the same lines with my cognitive subroutines theory, but I was a couple years too late (as well as lacking any sort of intellectual rigor). And I’ve already discussed how my localized generalities post was the same idea as the “invariant representations”, without the neuroscientific backing. So I’ve got some good ideas; I just need to develop them, and do the legwork to support them more fully. In my copious free time.
P.S. Speaking of which, wow, that was quite a lull in posting for me. It’s been crazy. My company is trying to finish up projects for two different clients before Thanksgiving, so I’m spending a lot of energy there. I had the conference in LA a few weekends ago. Last weekend was a company outing to Monterey. There’s all sorts of other social stuff going on. So on the few nights that I’ve had at home alone, I’ve been so exhausted that I have just collapsed comatose onto the couch to watch TV rather than blogging. But I wanted to get up the CellKey prototype pictures tonight, and then figured I should clear out the backlog of book reviews that has built up. I actually finished all of these books a few weeks ago (and then spent last week catching up on my backlog of The Economist), but they’ve just been sitting on my desk since then.