Tags

, , ,

Intra-Psychic Learning plays a crucial yet largely unacknowledged role in human intelligence. It will also play a critical role in so-called “artificial intelligence” or “the singularity.” In general, the paradigm most talked about in learning, whether by psychology professors or the general public, focuses on the role of external experiences. Famous examples include Pavlov’s dogs who exhibited classical conditioning. A bell was rung whenever food was presented and eventually the bell sound alone caused the dog to salivate. This works for humans as well. Just watch someone cut open a fresh lemon and you will find yourself puckering up and salivating! In operant conditioning, a rat learns, probably through a shaping process, that some behavior, say, pressing a lever, results in a reward such as receiving a food pellet. Eventually, the rat presses the lever. Both of these kinds of mechanisms are important and play a part in animal learning as well as human learning. Both kinds of learning are useful for AI as well. In humans (and to some extent in other animals as well), you do not have to “be in the loop” in order for learning to take place. You can *observe* another person getting a reward doing X and you might immediately try that behavior for yourself. Indeed, human beings take this one step further and can be induced to try (or not try) something based on what someone *says* about a behavior leading to a consequence. You don’t *have* to touch a hot stove and get burned or even watch someone else get burned by touching a hot stove in order to fear touching a hot stove. For most people most of the time, you can be told about hot stoves and that is enough. All these forms of learning focus on personal, observed, or bespoken information that actually exists about consequences in the real world.

However, there is another important way that we learn and it is based on checking intermediate results against each other without the need for any ground truth observation in the real world. I first mentioned this in my dissertation. I was studying human problem solving and fascinated by the observation that human chess players, who have excellent memories for real chess positions, would often examine one branch of a move tree, study another branch and then return to study the first branch again. This is not likely because they forgot. Instead, I believe that looking at the second branch taught them fundamental things about what was true for this particular chess position, and they then used that information to re-evaluate what they saw during their re-examination of the first portion of the game tree. Notice that in all of this thought process, they had not actually made a move in the real world and not seen their opponent’s actual response. They certainly did not yet get feedback about the ultimate outcome of the game.

In chess, as in many if not most endeavors in life, one may learn a great deal by examining things from various mental angles and comparing the results without waiting for actual feedback from the external world. Consider the case of a playwright writing a script. As they are writing, they are imagining the action, the facial expressions, the tone of voice. They are “checking” how the various characters react to what is being done and said. If something doesn’t “ring true” they will alter what they are writing. Of course, this process is not perfect and they may well make additional changes based on a reading and based on rehearsals. But many of the potential paths are already examined, selected and modified based on imagination alone.

Consider another interesting case that was extremely common through most of our evolutionary history and is still somewhat common today. A person walks through a physical environment. As they walk, they see before them a host of objects in a hypothesized set of physical relationships. In many cases, the information that is presented is extremely minimal at first. It is hard to tell whether that is a stranger over there or your cousin Bill. That looks like an oak tree, but maybe not. Is that a painting of some cedar trees on the side of that building or are those actual cedar trees over there? The brain is making a huge number of perceptual hypotheses about what these objects are and how they are arranged. As you move forward, you gain more detailed information. Now, you can clearly see that that is not your cousin Bill. That tree is definitely a sugar maple. Those are just well executed paintings of cedar trees and so on. You can use the difference in hypothesis weights between every two physical steps to update the weighting functions on all these perceptual hypotheses! You need not wait until you actually get verification that that is a maple tree. You do not wait until you reach the Bill-like stranger to make a modification in your weighting functions. In fact, you will probably pay little more attention to this figure as you approach. You already have enough information to learn. If, indeed, as you approach still more closely and Uncle Bill calls out to you —- making you suddenly realize you have prematurely concluded this was not Bill — you will again update your recognition function weightings. This may even come to consciousness and you may remark, “Uncle Bill! I hardly recognized you without your beard!”

This type of learning also plays an important part in improving sports performance. As a person improves their skill in golf, basketball, tennis, baseball, etc., they begin to anticipate earlier and earlier whether they have “executed” the move properly. An experienced tennis server, for example, generally knows long before their serve is called “out” that they have made an error. This process is not infallible, of course, but it is statistically better than chance, and for very skilled athletes it is much better than chance. You can see it when a slugger hits a home run and they take a skip step and watch the ball go out of the park. (There can be a downside to this facility of intra-psychic learning in sports under certain circumstances as explained in chapter 23 of The Winning Weekend Warrior). This means that the skilled athlete gets “feedback” from their own mental model of what they did critical seconds before a beginner does who must wait for feedback from the real world.

These kinds of phenomena are not limited to sight, or indeed, any one sense. You hear a very faint noise. You imagine it to be a cardinal singing. As you walk closer to the bird, you get a better signal and are more certain it is a cardinal. You can use the difference in certainty to internally reward those neuronal paths who were shouting “cardinal! cardinal!” And, you demote those neuronal paths who were shouting, “car backfire” or “firecracker” or “church bell.” If you get close enough to see the cardinal, you do even more internal tuning based on the inter-sensory verification. Similarly, if you walk toward what appears to be an uneven patch in the terrain, you imagine what you must do to compensate for that variation in the terrain. As you step on the uneven spot, your tactile and kinesthetic senses give you feedback about the terrain. You use this panoply of information from various senses to tune all of them.

While it is vital that, at the end of the day, we obtain feedback about actual consequences, a huge amount of human learning takes place simply by comparing what we think we know based on scant evidence to what we think we know based on slightly less scant evidence. I believe we are doing this continually within and across all our senses and that it actually accounts for the majority of our learning.

The Winning Weekend Warrior

Learning by modeling; in this case by modeling something in the real world.

Learning by modeling; in this case by modeling something in the real world.

Advertisements