“On Intelligence” by Jeff Hawkins

Summary

This book is surprisingly good in its ability to reach both the lay reader (for at least the first half) and the reader familiar with neuroscience. Articles since its publications provide much greater detail and are very useful for those interested in going deeper, but On Intelligence serves very well as an introduction to the concepts. The ideas expressed in On Intelligence are important both for scientific advancement and for philosophical consideration. While one could argue that perhaps there are other forms of intelligence or ways to produce intelligence, Hawkins does a good job in arguing what intelligence is in terms of mammalian brains and what the basic neocortical unit does. While Hawkins brings these ideas together in an orderly framework, he does give credit to the many neuroscientists responsible for the various components and underlying ideas that make it possible. These ideas as a whole until recently have not been sufficiently discussed in the neuroscience community in my opinion, and I believe they will aid (and in fact already have aided) greatly in advancing our understanding of the brain and creating real “artificial” intelligence that isn’t actually artificial at all.

The balance between addressing the expert and lay audiences did at ...

Book cover of On IntelligenceThis book is surprisingly good in its ability to reach both the lay reader (for at least the first half) and the reader familiar with neuroscience. Articles since its publications provide much greater detail and are very useful for those interested in going deeper, but On Intelligence serves very well as an introduction to the concepts. The ideas expressed in On Intelligence are important both for scientific advancement and for philosophical consideration. While one could argue that perhaps there are other forms of intelligence or ways to produce intelligence, Hawkins does a good job in arguing what intelligence is in terms of mammalian brains and what the basic neocortical unit does. While Hawkins brings these ideas together in an orderly framework, he does give credit to the many neuroscientists responsible for the various components and underlying ideas that make it possible. These ideas as a whole until recently have not been sufficiently discussed in the neuroscience community in my opinion, and I believe they will aid (and in fact already have aided) greatly in advancing our understanding of the brain and creating real “artificial” intelligence that isn’t actually artificial at all.

The balance between addressing the expert and lay audiences did at times leaving me wanting to see further caveats and explanations which often (though not always) did come later in the book. However the few issues I have with the book are relatively minor and do not affect the overall thesis or impact.

Here I will provide a bit more detail as to the concepts presented so you have a better idea of what you will find in the book. As many good theories do (though Hawkins correctly tries to avoid that work and use “framework” in its place), this one starts with looking at what it is our brains evidently do. They immediately recognize patterns, predict the immediate future without conscious attention (and raise stimuli to attention when they fail to fit the prediction), and represent things and concepts hierarchically in an invariant form at each level. Hawkins puts even more simply when he says that the brain is a memory retrieval device, in contrast to computers which do computations. I find the presentation of the dichotomy potentially misleading, but I do understand his point. As is often the case in the book, the truth of Hawkins’ statements resides in your definition of his words and to what system or components they apply.

Hawkins goes on to describe how the various anatomical layers in the neocortex enable this function, explaining that the input Layer 4 receives input from cortical regions with lower level representations, combine them in various ways into more complex representations given their temporally coincident occurrence/activation, then pass them to Layers 2 and 3 which generate sequences of those new higher order representations. He posits that Layer 2 specifically encodes a “name” for a temporal sequence encoded in Layer 3, and passes notice of their occurrence up to Layer 4 of a yet higher level of the cortex. Meanwhile they also send messages to Layers 5 and 6, which respectively pass information to motor areas to enable appropriate behavioral reaction and pass information back down to lower levels and thus enable the prediction capability. Thus, each cortical region, through the learning process, generates a more complex representation than the level below, which is independent of the specific detailed form of the input (invariant), creates temporal as well as spatial patterns, and then predicts the immediate future.

There is more detail, additional features to the framework, and a list of testable hypotheses for experimentalists to test. For instance, Hawkins discusses the role of the hippocampus in the framework. You may find as I did that some aspects of intelligence, human intelligence at the very least, is missing from the framework. It seems to me that top-down attention as well as lateral and non-hierarchical connectivity are left unexplored, and that these aspects of the brain are quite vital to our level of intelligence. However the framework as presented appears to be a potentially very powerful one. Indeed, Hawkins and colleagues, Dileep George in particular, have create an AI/machine learning version of this framework called Hierarchical Temporal Memory and have created successful businesses out of them. Now we have not only biological experiments to look forward to, but computer simulation, business, and technological experiments as well.

Leave a Reply

Skip to toolbar