
A book by Jeff Hawkins, from 2004, explaining his theory of brain and
intelligence.
Jeff takes the view that intelligence is the ability to (successfully)
predict. And the view that what brain does, on all levels, can basically be
summed up as memorizing patterns (and sequences of patterns) and providing
predictions based on those learned patterns.
More specifically, the brain is constantly extracting regularities from the
incoming data streams on multiple hierarchial levels: regularities in the
"low-level" sensory input, regularities in the occurrences of those low-level
regularities, etc. For example, it learns (roughly speaking) certain optical
patterns to form certain letters, letters to form certain words, words to form
certain sentences, sentences to form certain texts, etc. At the same time there
is also a constant flow of information in the opposite direction: the brain
associatively recalls, at all levels, based on previous experience, what should
be there and what should come next. I.e., it constantly makes predictions. And
if something does not match the predictions, an "alarm" is raised that should
lead to an (automatic) reconsideration of the situation and possibly to
learning something new. For example, if the brain assumes that the text that
the person started to read is some particular famous speech it knows, it
predicts what sentences, words and letters come next. If these predictions
fail, it may need to reconsider the initial assumption that the text is that
particular speech. Or if the prediction fails only on a single word, the
attention of the person is drawn to that word (which might turn out to be a
typo or maybe the word was incorrectly remembered by the person or just misread
in the first go at the moment).
And such continuous pattern extraction and prediction is what goes on in
almost all parts of brain and on all levels, and underlies everything from the
lowest sensory input and motor output to the highest levels of abstract
thinking. Jeff emphasizes that "Prediction is so pervasive that what we
"perceive" — that is, how the world appears to us — does not come solely from
our senses. What we perceive is a combination of what we sense and of our
brains' memory-derived predictions.", because what the brain predicts actually
affects what and how we sense and notice. Also, while when applied to incoming
sensory data the top-down flow has the role of predicting and drawing attention
to violations of prediction, the brain can also route the same top-down flow to
motor output, in which case it will be executed, if possible: "when your own
behavior is involved, your predictions not only precede sensation, they
determine sensation". In the given example of text processing, the particular
speech as such gets unfolded into sentences, and sentences into words, and
words into letters, and letters into hand movements for handwriting or hand
movements for typewriting or movements of the vocal apparatus for speaking
etc.
As of intelligence in general, Jeff claims that "Intelligence is measured by
the predictive ability of a hierarchical memory, not by humanlike behavior."
and that there is no reason to be afraid of intelligent machines, because "They
will not have personal ambition. They will not desire wealth, social
recognition, or sensual gratification. They will not have appetites,
addictions, or mood disorders. Intelligent machines will not have anything
resembling human emotion unless we painstakingly design them to."
I find the central point of the book very interesting and quite likely to
have a strong relevance to how the brain really works (not necessarily in
details, as Jeff himself points out, but at least in some of the general
principles). However, I am not so sure about the reliability of those parts
that touch upon artificial intelligence more generally, outside his
memory-prediction framework. I cannot really disprove anything, but I
occasionally got the feeling that some things are not fully correct. For
example the aforementioned belief that intelligent machines will be just
emotionless pattern detectors and predictors unless we painstakingly embed the
emotions — some alternative opinions say that such emotions and drives are
actually crucial for developing interesting higher level AI (then again, maybe
it's partly about the difference of goals — whether we should create
emotionless pattern detectors-predictors or systems that behave
interestingly). Also, I get the impression that when talking about AI Jeff
seems to equate it only with the early classical logic-based AI, and when
talking of "robots" he seems to think of inflexible systems without much any
feedback... (which isn't exactly the case in general).
Also, even though Jeff emphasizes that he is not interested in
building humans, but in understanding intelligence and building intelligent
machines, he categorically claims that "We have to extract intelligence from
within the brain. No other road will get us there.". I agree that brain is the
best example for us to build on, but by no means should we be so categorical
and exclude other roads to high-level machine intelligence. Jeff points out
early in the book that it was his strong intuition that made him rather sure
that the Artificial Intelligence approach will fail to create programs that do
what humans can do and will also fail to teach us what is intelligence in
general. While I have actually had the very same intuition for a long time as
well, and am thus very supportive to nonclassical approaches to AI, I
nevertheless exercise caution about getting channeled into the other extreme,
and I find it slightly alarming that Jeff seems (in my humble opinion) to fall
into the very trap that he, ironically, cautions against just 20 pages later:
"However, looking across the history of science, we see our intuition is often
the biggest obstacle to discovering the truth".
But these small complaints of mine are mainly about the peripheral / general
thoughts in the book, not about the core point which I find very interesting.
So, overall I'd still say that the book "On Intelligence" is definitely worth
reading for anybody interested in the workings of the brain and (but not
necessarily) in how to create thinking machines.
More info about the book at Amazon: http://www.amazon.com/dp/0805078533
Last comments