On Intelligence - take-away lessons

Some take-away lessons from On Intelligence by Jeff Hawkins...

The book concentrates on the neocortext (aka the cortex), a cauliflower-like sheet of tissue on the outer surface of the brain. "Almost everything we think of as intelligence — perception, language, mathematics, art, music, and planning — occurs here. your neocortex is reading this weblog entry.

The average human neocortex, spread out, would be about the size of an unfolded dinner napkin, built of 6 layers, in total about the thickness of a 6 business cards. Anatomists estimate the typical human neocortex to contain around 30 billion neurons (perhaps less, perhaps more). These cells contain all of your memories, knowledge, skills, and life experience. They are "you".

The hippocampus is the "uppermost" level of the neocortex. That's pretty neat. The hippocampus creates longterm memories (which are then stored in the neocortext).

Jeff agrees with a theory proposed by Vernon Mountcastle of Johns Hopkins University on 1978. The theory proposes "An Organizing Principle for Cerebral Function" and suggests that the neocortext uses the same computational algorithm for everything it does — that is, touch, hearing, smell, speaking, sight — are all handled in the same basic way. Your senses may not be the same, but the way the neocortex processes the signals it receives is the same, regardless of where those signals come from. This is not a popular theory among scientists and engineers, but by the end of the book I thought Jeff had made an excellent case for it! He refers to Moutcastle's paper as the "Rosetta stone of neuroscience".

Brains rely on patterns. I believe that. I've always understood that people see patterns, look for patterns, find patterns in everything.

All the information that enters your mind comes in as spatial and temporal patterns on the axons.... Spatial patterns are coincident in time; they are created when multiple receptors in the same sense organ are stimulated simultaneously. Temporal patterns [are] constantly changing over time.

All of our senses use both spatial and temporal patterns (even if we don't notice that, consciously).

Brains do not work like computers. Neurons are quite slow in comparison to the components of a computer. A ty[pical neuron can collect information, send it to other neurons, and reset itself, in about 5 milliseconds (i.e. 200 times per second). A basic current computer operation is five million times faster. Yet, a human brain can perform certain significant tasks in much less than a second. For example, a human can look at a photograph and, in half a second or less, determine if there is, for example, a cat in the photo. Your job would be to press a button if there is a cat. This is a task that is difficult or impossible for a modern computer to perform today, yet a human can do it reliably in half a second or less.

... neurons are slow, so in that half of a second, the information entering your brain can only traverse a chain one hundred neurons long. That is, the brain "computes" solutions to problems like this in one hundred steps or fewer, regardless of how many total neurons might be involved. From the time the light enters your eye to the time you press the button, a chain no longer than one hundred neurons could be involved. ... One hundred computer instructions are barely enough to move a single character on the computer's display, let alone do something interesting.

Instead of computing answers to problems [like a computer], the neocortext uses stored memories to solve problems and produce behaviour. The neocortex

• stores sequences of patterns
• recalls patterns auto-associatively
• stores patterns in an invariant form
• stores patterns in a hierarchy

Jeff provides a thought experiment. Assume that while you are out of the house, I modify your front door in some way. Perhaps I paint it, change the latch, or replace a wooden door with a metal door or a solid door with a hollow-core model. When you return home, it will take you a very short time to realize that something has changed. There is only one way to interpret your reaction to the altered door.

At every moment, your brain makes low-level sensory predictions about what it expects to see, hear, and feel. "Prediction means that the neurons involved in sensing... become active in advance of them actually receiving sensory input". When the input arrives, it is compared with what was expected.

Prediction is... the primary function of the neocortex. If we want to understand what intelligence is, what creativity is, how your brain works, and how to build intelligent machines, we must understand the nature of these predictions and how the cortex makes them. Even behavior is best understood as a by-product of prediction.

The hierarchical nature of the world is mirrored by the hierarchical (nested) structure of the neocortex. All objects in the world are composed of subobjects that consistently occur together. This one really hits home for me.

The design of the neocortex and the method by which it learns naturally discover the hierarchical relationships in the world. ... The cortex has a clever learning algorithm that naturally finds whatever hierarchical structure exists and captures it.

... Predictability is the very definition of reality. If a region of the cortex finds it can reliably and predictably move among these input patterns... and can predict them accurately as they unfold in time... the brain interprets these as having a causal relationship. The odds of numerous input patterns occurring in the same relation over and over again by sheer coincidence are vanishingly small... So reliable predictability is an ironclad way of knowing that different events in the world are physically tied together.

...Therefore, the brain can be said to store sequences of sequences. Each region of the cortex learns sequences, develops what [Hawkins] will call "names" for the sequences it knows, and passes these names to the next regions higher in the cortical hierarchy.

...By collapsing predictable sequences into "named objects" at each region in our [cortical] hierarchy, we achieve more and more stability the higher we go. This creates invariant representation. ... The opposite effect happens as a pattern moves back down the hierarchy: stable patterns get "unfolded" into sequences.

...in a complimentary bit of efficiency, representations of simple objects at the bottom of the hierarchy can be reused over and over for different high-level sequences.

Wow. Good stuff.