AI Influences Brain Science

Just as AI benefited from the example of neuroscience, ideas flowed in the other direction.

Brain Architecture

As early as the 1960s Blum’s colleagues concluded that there must be a way for neurons to inhibit the firing of other neurons, reasoning from a computational viewpoint. Neurobiologists subsequently discovered it was true.

While neurobiologists worked upwards from brain studies using fMRI (functional Magnetic Resonance Imaging), more progress was made by psychologists working from human behavior downward toward the neurons. The cognitive psychology of Newell and Simon had a distinctly computational flavor. Simon’s theory of humans’ bounded rationality arose from his recognition that every computation device has limits.

Possibly inspired by the small CPUs of computers, psychologists have shown that we have two kinds of memory: short-term memory, which can hold at most about seven items—such as digits or letters—for at most thirty seconds, and long-term memory, which holds everything else permanently. They call the items chunks but don’t know precisely how they’re represented.

More recently Nobelist Daniel Kahneman described experiments that suggested the brain has two thinking systems. System 1 consists of our built-in, subconscious processes and does 99 percent of the thinking quickly. System 2 applies a slow, rational process, sometimes used to justify the choices of System 1.

Maybe System 1 is in long term memory; System 2, in short term memory.

The indefatigable Blums are filling in the story with a more detailed sketch for the brain. First, they say that our consciousness is the chunk in short-term memory, which resides at the root of two branching structures that connect with the long-term memory—which we’ll call the subconscious. The subconscious holds millions of other chunks that sense the environment, drive our bodies, and hold long-term memories. Our consciousness receives one chunk at a time from the subconscious, holding it in short-term memory. Sometimes the consciousness poses questions to the subconscious in the form of a chunk, which is transmitted simultaneously via one of the branching tree structures to all the parts of the subconscious, which operate in parallel. Their various answers, also in the form of chunks, are transmitted back in the second tree, ending at the consciousness. As they converge, they compete in a kind of tournament based on their vehemence so that only one chunk arrives at the consciousness. This is reminiscent of Selfridge’s pandemonium model for recognizing characters.

Thus, thinking is a sort of conversation among all the processes moderated by the consciousness. Asked, “What Greek estimated the size of the Earth?” you might say “I know, but it’s on the tip of my tongue.” But the question has been transmitted to your subconscious processes that begin churning away. A few minutes later, the name “Eratosthenes” pops into your head. It presumably won contests against “Plato”, “Euclid”, and the others as it moved through the tree leading to the short-term memory.

This theory is obviously incomplete and is a work in progress, like all theories.

It seems plausible that System 1, the subconscious, is what we call muscle memory. There are fMRI studies that show many actions we take with muscle memory are made before our consciousness even notices the need for them. Every athlete understands that they first think about a new move and then practice it consciously until it’s in their muscle memory. It is also well-known that practice of any kind reduces the time to perform it according to a power law formula like T = F/N^C where F is how long it takes the first time, N is the number of times it is practiced, and C is a constant ranging between 1 and 2.

To explain the speedup, Newell suggested that conscious actions, e.g. “swing left arm forward with right foot, then swing right arm forward with left foot,” are carried out sequentially, but later get bundled up as a single chunk somehow and stored in long-term memory under something called “run.” Once they are subconscious and the chunk “run” is sent to the long-term memory, they can be performed faster because some can be done in parallel, and they can be sequenced without communication back and forth with the consciousness. This process cascades. The conscious action “run to the right while swinging the racquet back, then swing forward to hit the ball” bundles up those actions into a chunk that can execute faster in long-term memory.

For example, suppose some activity is composed of a hierarchy of chunks looking like this:

Assume each dot takes a millisecond to perform. Now here is the way the hierarchy might become imbedded in the subconscious as it is practiced.

As the lower actions are made subconscious and are done in parallel, so their dots merge, and the time decreases to about ½, ¼, and ⅛ in this case.