''On Intelligence'' by Jeff Hawkins.
Amazon review (5 out of 5 stars).
September 19th 2010.
- Simon Laub
-------------------------------
The ability to make predictions about the future is the crux of intelligence!
And Jeff Hawkins book ''On Intelligence'' presents some brilliant ideas
on how the brain might be doing this.
Sure, some might say that because the brain is so complicated, we will
never really understand how it works.
But according to Hawkins, complexity is a symptom of confusion.
Indeed, we need some core ideas that can help us make sense of the whole thing.
In Hawkins book, the core idea is seeing the brain as a memory-prediction system.
A memory system, that store experiences in a way that reflects the true structure
of the world. A system that remembers sequences of
events and makes predictions based on these memories.
According to Hawkins, such a system is the basis of human intelligence, perception,
creativity, thoughts and even consciousness.
=====================
In Hawkins presentation, the seat of intelligence is the neocortex. Even though it
has a great number of abilities and powerful flexibility, the neocortex
is surprisingly regular in its structural details.
The different parts of the neocortex, whether they are responsible for
vision, hearting, touch or language, all work on the same same principle.
The key to understanding the neocortex is understanding these
common principles and, in particular, its hierarchical structure.
According to Wikipedia: The neurons of the neocortex are arranged in vertical structures called neocortical columns. These are patches of the neocortex with a diameter of about 0.5 mm (and a depth of 2 mm). Each column typically responds to a sensory stimulus representing a certain body part or region of sound or vision. These columns are similar, and can be thought of as the basic repeating functional units of the neocortex. In humans, the neocortex consists of about a half-million of these columns, each of which contains approximately 60,000 neurons (However, because of inhomogeneity in columnar borders outside the visual cortex, further studies are needed before accepting the columnar cortex theory as fact).
When Francis Crick wrote the book, ''The Astonishing Hypothesis'' -
the astonishing hypothesis was simply that the mind is the creation
of the cells in the brain -There is nothing else, only neurons in a
dance of information. Hawkins agrees. And for him, understanding
intelligence, is a question of understanding how the thirty billion neurons
of the neocortex works together.
Other people might tell us, that you cannot possibly understand the neocortex without understanding
brain region x,y or z, because they are highly interconnected.
Hawkins doesn't disagree, the brain consist of many regions
and most of them are critical to being human
(Yet, sometimes people can live without certain
brain regions, see my ''Consciousness beyond life'' text).
But Hawkins is not interested in building humans - he wants to understand
intelligence (and build intelligent machines).
And for that purpose he thinks it will be enough to study the neocortex.
A part of the brain strictly related to intelligence (Human
sexual urges, hunger, muscles, emotions etc. are all very interesting. But, according to
Hawkins, not relevant in (t)his study of the neocortex).
Sure, anatomists have for decades recognized that the cortex looks similar everywhere.
But instead of asking what that could mean,
they went on to look for differences. And they did find
differences. If you look closely enough you'll find them.
Regions of the cortex vary in thickness, cell density, etc.
But according to Hawkins, it is the similarity that
is surprising and interesting, much more so than
the differences (I asked a doctor friend of mine about this. She believes that neurons might look the same in
different parts of the cortex, but probably there will be many differences in brain chemistry etc).
=====================
Hawkins is especially inspired by Vernon Mountcastles 1978 paper: ''An organizing principle for Cerebral
Funcion''. In the paper Mountcastle points out that the neocortex is
remarkably uniform in appearance and structure. The regions of
cortex that handle auditory input look like the regions that
handle touch, which look like the regions that control muscles,
which looks like Broca's language region, which look
like practically every other region of the cortex.
Mountcastle suggest that since these regions all look the same,
perhaps they are all actually performing the same basic
operation! He proposes that the cortex uses the same computation tool
to accomplish everything it does.
Sure, input is flowing in from the five senses:
Sight, hearing, touch, smell and taste
(And actually there is a lot more. Vision is really motion, color and luminance.
And touch has pressure, temperature, pain and vibration. And we have sensors that
tell us about our bodies position, the proprioceptive system.
And systems which gives us our sense of balance, the vestibular system in the inner
ear etc. ).
But, the inputs to your cortex are all basicly alike.
You hear sounds, see light, and feel pressure,
but inside your brain it is just signals. There isn't
any fundamental difference between these signals.
Its patterns. Spatial and temporal patterns. Where spatial patterns
are coincident patterns in time. Patterns where multiple receptors in
the same sense organ are simulated simultaneously.
Temporal patterns are the things that change over time.
Now, Hawkins idea is that
there is a single powerful (learning) algorithm implemented by every
region of the cortex. If you connect regions of
the cortex together in a suitable hierarchy and provide a stream of input
it will learn about its environment.
Learning how this cortical algorithm works will be absolutely essential for any real
progress in understanding (human) intelligence.
=====================
The job of any cortical region is to find out how its inputs are related,
to memorize the sequence of correlations between them,
and to use this memory to predict how the inputs will behave in the future.
The brain doesn't ''compute'' answers to problems.
It retrieves the answers from memory. And that is
why the brain can be so fast, even though neurons really aren't that fast.
It only takes a
few steps to retrieve something from memory.
Slow neurons are not only fast enough to do this,
but they constitute the memory themselves. The
entire cortex is a memory system. It isn't a computer at
all.
Our brains use stored memories to constantly make predictions
about everything we see, feel and hear.
Prediction is so pervasive that what we perceive -
that is how the world appears to us - does not
solely come from our senses. What we perceive is a combination
of what we sense and of our brains memory derived predictions.
Looking around a room, the brain uses memories to form predictions
about what we expect to experience, before we actually experience it.
Is the window rectangular and the walls vertical? Yes. Is sunlight coming
from the correct direction for the time of day? Yes.
When some visual patterns comes in that is not memorized in the context,
a prediction is violated. And attention is drawn to the error.
And thats why we can be intelligent - just sitting in a corner of a room,
observing. We will constantly be making lots of intelligent predictions on what is going
to happen next!
Indeed, neuroanatomists have known for a long time that the brain is saturated with
feedback connections. E.g. in the circuit between the
neocortex and the thalamus, connections going backward (towards the input)
exceed the connections going forward by almost a factor of ten.
That is, for every fiber feeding information forward in the neocortex,
there are ten fibers feeding information back towards the senses.
Indeed, feedback dominates most connections throughout the neocortex. Prediction is just as important as actually sensing something.
=====================
The neocortex memory-prediction system is nothing like the memory system of
a computer though!
Hawkins states it like this: a) Computer memory does not (normally) store sequences
of patterns. In contrast, the cortex does store sequences
automatically. b)The cortex is an auto-associative system. One that can
recall complete patterns, when only given partial or distorted
inputs.
Working both on spatial and temporal patterns - Thoughts and memories are associatively linked.
So inputs auto-associatively link to what follow next.
This gives the chains of memories we normally call
thoughts. c) Memories are stored in a form that captures the essence of relationships,
not specific details of the memory..... When you see,
feel or hear something, the cortex takes the detailed,
highly specific input and converts it into an invariant form.
It is the invariant form that is stored in memory. And
new input patterns gets compared to the invariant form.
Memory storage, memory recall, and memory
recognition occur at the level of invariant forms.
There is no equivalent concept in computers....
The three properties of cortical memory - a) storing sequences,
b) auto-associative recall, and c) invariant representations
are (according to Hawkins) the necessary ingredients to predict the future based on memories
of the past....
With such a ''memory'' - we can think about the world, and move around in the world,
and make predictions about the future, because the cortex has
built a model of the world.
The real worlds nested, hierarchical structure is mirrowed
by the nested structure of the cortexs memories.
Indeed, one of the really clever parts of the cortex is its ability
to capture whatever hierarchical structure that exists
and store it (And obviously, when structure is absent, we are thrown
into confusion).
In more detail: A column of the cortex will handle a certain activity and
pass on a stable pattern (''the name of the activity'') to its higher region.
If the column can't predict what is going on it will pass the information
further up the hierarchy.
When every higher region says ''Don't know what this is.
lets skip this to a higher up region'' you eventually get to the top of
the cortical pyramid, where information that can't be understood
by previous experience ends up. The new and unexpected input.
This enters the Hippocampus and is stored there (The neocortex appeared on the evolutionary scene sandwiched between
the Hippocampus and the rest of the brain). But it won't be stored there forever, either it will
later be transferred back into the cortex or it will be lost forever.
That is - when we learn something - Representations move down the hierarchy.
I.e. through repeated training, the cortex will relearn
sequences in hierarchically lower regions of the cortex.
In the end we will be experts:
Seeing patterns and higher order structures is what makes us experienced (The really good expert can see patterns beyond what anyone else can see).
=====================
Hawkins agrees that it is difficult to imagine, how a simple but numerically vast
system can create our consciousness, our languages etc. - But he suggest that
it is because our intuitive sense of the capacity of the cortex
and the power of this hierarchical structure is inadequate.
He sees it as a fact, that evolution discovered that if you attached a hierarchical
memory system to the senses, the memory system would model the world
and predict the future.
Now, the real interesting thing is to figure out precisely, how this neocotical algorithm
works, and start building simulations of it.....
According to Hawkins: ''The question of intelligence is the last great terrestrial frontier of science.
Everyone has a brain. You are your brain. If you want to understand why you feel
the way you feel, how you perceive the world, why you make mistakes,
how you are able to be creative, why music and art are inspiring, indeed what it is
to be human, then you need to understand the brain''.
And certainly, Hawkins thinks that computers will some day be programmed
with ''cortical algorithms'': Sure, a computer could model all the neurons and their connections, and
if it did, there would be nothing to distinguish the ''intelligence''
of the brain from the ''intelligence'' of such a computer simulation.
But such (machine) intelligences will be nothing like human minds.
Without the ''old brain'' parts, such machines will not have personal ambition,
they will not have desires (i.e. all the things from our ''old'' limbic brain: Fear, paranoia,
desire .... all the stuff that makes us human....).
And the speed difference between organic and silicon based minds will be of great
consequence. I.e. Intelligent machines will be able to
think millions of times faster than human minds. So, such articial minds will be able to read and understand libraries of
books and study complex bodies of data - in mere minutes.