A Mathematical Theory of Human Intelligence (Neocortex)

Jeff Hawkins' Hierarchical Temporal Memory

Jeff Hawkins' Hierarchical Temporal Memory theory of intelligence takes the human neocortex as its starting-point. The human neocortex, the "new" cortex, is our evolutionarily most advanced and recent development that makes humans human. Anatomically, the neocortex makes up around 70% of the brain volume, and is wrapped around the more primitive structures of the brain like a large sheet that is folded, which gives rise to the familiar, characteristic look of the brain with folds everywhere on the surface. The neocortex is structurally self-similar, and from a material perspective it makes sense that the sheet does similar kinds of computations everywhere. This is also plausible from an evolutionary perspective, since evolution tends to stick with what works, copying rather than inventing all anew.

By studying the way in which the neocortex is processing information, the expectation is that we can unravel something about the nature of our general intelligence, and the theoretical hope is that there is some theoretical unification that makes a neat explanation possible. Jeff Hawkins has proposed a general theory of how the neocortex works, his Hierarchical Temporal Memory theory of intelligence. There have been earlier theories to the same note by Fukuyama and other more recent theories such as HMAX that explain things a little differently. The ultimate motivation seems to be the same, and the underlying hope seems also to be the same.

Recent successful startup Vicarious was started by a Jeff Hawkins' co-author and co-founder of Numenta, Dileep George, and is currently funded by, among others, Elon Musk, Mark Zuckerberg, Peter Thiel. In other words, big money is betting on this approach leading somewhere within the next decade or so.

I think Jeff Hawkins' is very good at explaining the general gist of the approach and the progress, detailed in the video above and his book "On Intelligence." While this approach differs from Friston, Schmidhuber and Hutter, I think ultimately all of these things will converge in a common framework.