Challenged my thinking
This book provides some very stimulating insights into how human being go about the process of thinking and how the brain functions. It helps you understand why things like Artificial Inteliigence are no where near matching the marvel that is the human brain.
To get the most of this book you will need to sit down and concentrate because there is plenty in here to digest. It is not something I'd call bed time reading. If you are interested in learning how the human brain functions and what makes it do what it does then this book is for you.
Personally, it has changed the way that I look at many things about brain functions. I also reckon it is going to help me better understand my own brain and get more from it. In short, a really worthwhile read.
The best of a bad bunch
This is not a comprehensive review, just some thoughts on the subject I consider my specialty (been working on it for 20+ years). I give the book 5 stars not because it's perfect, but because there's no better generalization on the subject. It's amazing how little of meaning is written on the most important phenomenon in the universe: intelligence. Hawkins has a good intuition, but he doesn't follow through to consistently formalize it. This is an excerpt from my knol:
The notion of cognitive hierarchy is a key, but Hawkins seems to be confused as to what it selects for: generality or novelty. He nominates both, apparently not realizing that they are mutually exclusive: the scope of discoverable generality is the span of inputs a pattern have searched through, & the longer it searches the less novel it gets. I think generality & novelty are selected only to the extent that each contributes to predictive correspondence:
The value of generality is obvious: predictive power of a pattern *is* its projected generality.
Novelty is more ambiguous because it has two different aspects: proximity & change.
Recent inputs are relatively more predictive than the old ones by the virtue of their proximity to future inputs. Thus, proximity should determine the order of search within a level of generality. It can't be a criterion for hierarchical selection: elevation expands search, reducing proximity among comparands. Change, on the other hand, has a contrast effect: it may interrupt a pattern. In that case, change has a negative value: it's important only to the extent that it cancels positive predictive value of interrupted pattern. Change within noise has no value because it doesn't interrupt any pattern.
However, human curiosity is not purely cognitive, it's biased toward the survival utility of information. Proximate & changing objects are more likely to affect the subject in a short term, thus attracting far more attention than they would for their contribution to predictive power alone.
Hawkins' HTM model is less ambiguous than his "On Intelligence", - it selects for recurrence only. However, the comparison & subsequent selection there is very coarse-grained: he doesn't quantify match per variable, or even per 1D sequence of inputs. He starts by comparing 2D frames: the level of complexity that leads to combinatorial explosion almost immediately. I'd start from the beginning by comparing pixels: the limit of resolution, & quantify the degree of match as a distinct variable. It also necessary to record & compare explicit coordinates & derivatives, which he ignores as meaningless. I believe HTM & similar ANN models don't scale largely because their selection process isn't incremental enough.
1more hour I won't get back
I'll admit I only made it a third of the way through the book, but up until that point, his assumptions first pop out of bad math. Then he builds wacktastic theories on top of his assumptions. I'm glad those theories opened the creativity for the positive reviewers, but for me, it wasted an hour of my life.
Innovative Model of the Human Cortex
Hawkins has created an electrical engineer's view of how the human cortex creates patterns from sensory input, stores the information, and uses patterns predictively. The book is an easy read and very entertaining. A recently published peer-reviewed paper by the author (2009) contains the messy details and should cause a stir in the artificial intelligence community if nothing else.
A New Way to Think About Intelligence
After listening to a podcast interview with Jeff Hawkins, I picked up this book because although I'm by no stretch of the imagination an AI expert, Hawkins' arguments regarding the failures of AI research over the decades made sense, and so I thought it would be interesting to learn more about his particular take on the topic. Hawkins certainly delivers in that regards, offering a theoretical framework for his conclusion that the brain is essentially a highly organized prediction machine which manages to outperform even the most powerful of today's computers despite the brain being woefully slow comparatively.
To be clear, this is not a book about "artificial intelligence", but rather focuses on how the human brain operates. The authors devote a mere 30 pages (Chapter 8: The Future of Intelligence) to a specific discussion of how Hawkins' theory might apply to technology development. Do not construe this as being a shortcoming, because Hawkins' intent (as I understand it) is to right his perceived listing of the AI ship by first rethinking the concept of intelligence before applying these theories to silicon.
If you've any interest in the science behind what may one day make the machines of "I, Robot" a reality, consider reading this book.
(118 reviews)
