April 24, 2015 @ 1:06 pm
While I was at Cold Spring Harbor Labs a few weeks ago, I picked up a copy of On Intelligence, by Jeff Hawkins, formerly of Palm Computing (remember the Palm Pilot?) and a science writer named Sandra Blakeslee. It's a ten-year-old book, which hasn't made the spash it deserves. At the time I seem to remember Hawkins being dismissed as a Silicon Valley pretender, a dilettante, and I never got around to reading it. Fortunately, he's on the CSHL board, and they keep copies of it in their bookstore.
The most basic idea of the book is actually quite similar to Darwin's insight on how to organize biological diversity. Darwin realized that each species was not a separate thing. Every species on the planet is related. Once we take time into account, we can see this organization, the Tree of Life into which each individual species can be plugged. That reorganization of biological diversity has been incredibly valuable, because it allows us to generalize responsibly from one species to another based on how far apart they are in the tree. Insects like fruit flies are a great model system for genetic disease because they share all those mechanisms of sexual reproduction with us. Asexual bacteria are good models of more basic processes like energy metabolism, but less good for other things like cancer.
Hawkins had a similar insight. The 52 Brodmann areas of the human cerebral cortex are not separate computational modules, separately evolved to perform specific computations, the way generations of neurologists and cognitive scientists have imagined. Those cortical areas are all related. They all do the same thing; they just do it on different inputs. Visual cortex performs the same operations on the signals from the optic nerves that auditory cortex performs on the signals from the ears. That's a powerful idea, for the same reason that Darwin's idea was powerful. If it's true (and I suspect that it is true), we'll be able to leverage discoveries from one field of neuroscience to others, and we can learn which discoveries will generalize and which discoveries will not. That's big.
If we think about the biological evolution of the brain as an organ, this makes perfect sense. Reptiles don't have a cortex, and neither do birds. They have a 3-layered hippocampus, which is similar, and probably the ancestor of the cortex. As far as we know, the 6-layered cortex arose one time in the history of mammals, not dozens of times. New cortical areas are built out of more or less the same stem cells as old areas. Simplicity suggests that they should do the same thing.
Hawkins goes further. He says that our larger model of the cortex as a computer is wrong. The cortex does not compute equations, as most model-builders assume; it's not a processor. The cortex is a memory system. It computes by analogy to its stored memories. The goal of the cortex is to predict its next input, much as the goal of the much older cerebellum is reduce the error signal in movement.
Try touching your nose. The error is the difference between where your finger currently is and where you want your finger to be. If your cerebellum is healthy (and you haven't been drinking alcohol), that error signal will be very small, because at the end of the movement and at every point along that trajectory your finger will be exactly where you expect it to be. If you put on a pair of prism glasses, as I do with my Governor's School students every summer, and shift the visual input by 10 degrees to the right, you've created a big error signal, which the cerebellum will eliminate over the course of a few minutes. Your initial throw of a football will be 10 degrees off. Each throw becomes slightly more accurate until things feel normal. When you take the glasses off, the error signal returns, but in the opposite direction, so that the cerebellum has to zero out that signal, too.
Hawkins believes that the cerebral cortex does the same thing, not in 3 spatial dimensions but in the high-dimensional spaces of similarity. As we move up the cortical hierarchy those spaces become more and more abstract, more and more compressed. For instance, the 3 spatial dimensions might be compressed into a single abstract movement profile, so that I can recognize the way my wife moves across a crowded room.
OK, interesting – but what does this mean for teaching? It means that, contrary to our fervent hopes, we can't simply download that abstract representation into the brains of our students. It has to be built up through repetetive experience. We may be able to catalyze that process by offering good examples and guiding the generalization, but we can't simply present a rule or a concept or a generalization and drop the mike the way we imagine that we can. That's a hard truth to accept. (I'm right now listening to a student complain about “trick questions” at the table behind me at Starbucks.)
What's worse, students are really adept at repeating those rules back to us, so it can look as though they've got it when they really don't. In fact, even they believe that they've got it, because they can repeat the rule back. It feels as though they have zeroed the error signal, because we're all measuring the wrong signal, the easy convenient signal of verbal match rather than the important, buried signal of analogy match. Measuring analogy match requires presenting novel examples and measuring a much more complex error signal. The Venn diagram of new information overlapping with old experience becomes very complex, like Luisa Hiller's diagrams of the pangenome of related bacterial species.
The dirty secret? Many teachers don't understand the concepts themselves, so they aren't equipped to do more than the rote verbal lecture, even if they wanted to. So we have a lot of work to do. I think Hawkins's book is a good place to start.Share | Comments