Natural Forgetting Algorithm

Last modified: 2008-08-12 13:05:03 UTC

A theory concerning natural learning is presented. This theory places emphasis not on how the brain learns per se, but rather, on how the brain selectively forgets, wherein it is the forgetting that leaves the learning in place by default.

Unbelievable advances have been made recently in our understanding of neural nets. Studies in massively parallel distributed processing, using error backpropagation algorithms to weight the connections among nodes, have laid the foundation for the consideration of the possibility that the most sophisticated functions in the human brain might one day become explainable in mechanistic terms.

Nevertheless, error backprop is fundamentally unrealistic, as it requires that the correct answer already be known, and then the connections in the MPDP network can be weighted properly to yield the correct answer.

The real question is how the brain can learn directly from Mother Nature, without already knowing the correct answer internally, and perhaps without Mother Nature having a way of poking error backprop signals directly into specific neurons to affect the appropriate changes.

It's possible that we should stop focusing on how the brain learns, and start taking a close look at the circumstances under which the brain forgets, and consider the possibility that it is actually selective forgetting that is the business end of natural learning.

Consider the well-known phenomenon that people who have been through a traumatic event, such as a car accident, frequently experience mild retrograde amnesia with regards to the events immediately preceding the event. From a Functionalist standpoint we might consider this to be odd, since if the brain's job is to prevent injury to the body, then the brain would naturally want as much information as possible about the circumstances that led to injury, so as to avoid similar circumstances in the future.

There are a number of perfectly plausible explanations for the amnesia, wherein the cognitive effect is an artifact of physiological processes, and serves no functional purpose. For example, the traumatic event caused the release of adrenalin, which, in addition to facilitating enhanced muscular responses, also constricted blood capillaries. This constriction serves the primitive purpose of reducing blood loss in the event of injury, but then might also have the unintended side-effect of depriving the brain of the nutrients that it needs for the transformation of short-term memories to long-term memories. Hence the retrograde amnesia might have no functional significance.

On the other hand, what if it did? What if Mother Nature is results oriented, and does not care how those results are achieved? What if the retrograde amnesia is an intended side-effect? Perhaps Mother Nature is saying, "That neural activity pattern led to injury, so just forget the whole thing."

This establishes how we might forget thoughts that led to bad consequences. But how do we remember thoughts that were successful?

The proposed answer is that we remember these thoughts by default. As a matter of fact, this theory states that we remember everything by default, and it is merely those activity patterns that were immediately succeeded by traumatic events that we forget.

It's also possible that the event doesn't actually have to be traumatic, to the point that primitive physiological processes were triggered, in order to disrupt the transition of short-term memories to long-term memories.

A study was once done that tested the ability to remember sequences of numbers under different circumstances. The researcher found that if a number was shown to a test subject, and then there was a 30 second period of silence, the test subject was able to write down the number. But if a bell sounded partway through the 30 second period, the retention rate was substantially lower. This extremely simple test may have been demonstrating something quite central, not just with regards to the study of memory, but to the study of learning in a far more general sense.

It's possible that any activity pattern that lingers will be retained. And it's possible that any activity pattern that is disrupted, by whatever means, is less likely to be retained. And it's also possible that this would have enormous functional significance.

In the most basic sense, it is the job of the brain to protect the body from injury. For this purpose, the body has been wired up with all kinds of sensors to detect biological insult. The brain is then left to engage in whatever activity pattern it wants, so long as it does not result in one of the sensors being triggered. If that happens, the current activity pattern is disrupted, and therefore is not migrated to long-term memory.

Let's consider how this theory might work in the real world. For example, consider a person walking through the woods. For the sake of simplicity, let's think of this person as having no existing knowledge of any kind, with a perfectly empty brain, but with a generic ability to learn, and with some reason to be walking.

It will only be a matter of time before the person will run into a tree. This will result in biological insult to one or more parts of the body. This means that the activity pattern involving seeing a tree take up more and more field of view, combined with the somatosensory feedback generated by continuing to walk towards it, will not transition to long-term memory.

After bouncing off of the tree and walking away in another direction, he/she will run into another tree, with the same effect. This process could be repeated an indeterminate number of times.

Eventually, as the person approaches the next tree, and sees the tree taking up more and more of his/her field of view, by chance the person might turn slightly, and avoid the tree. This activity pattern will then be retained, because biological insult did not result.

The next time the person approaches a tree, he/she will then turn slightly, and thereby avoid hitting it. This is not because he/she knows that this is the best thing to do, or that pain would result if he/she did not. The person will exhibit this behavior simply because this is the behavior that the person knows the best. The neurons involved in this activity pattern fire more robustly than any other, and this is the response.

Note that this constitutes one-trial learning. Within this theory, all that is required for learning to occur is for the activity pattern to somehow become active, and for biological insult to not be the result. In essence, this theory states that all trials result in learning except the ones that failed.

All that would be required for this algorithm to be instantiated in vivo would be for there to be some sort of delay in the mechanisms that form long-term memories. In other words, whatever you're thinking right now, if you're still thinking it two minutes from now, your current thoughts can start the migration to long-term memory.

There are any number of mechanisms that could be proposed for this, and many researchers believe that there are a number of different types of long-term memory, all with different biological instantiations, and all operating on different time frames. Dilation of blood capillaries, build-up of nutrients within the cell bodies because of demand from the axons, increased amounts of nutrients within the axons because of demand from the synapses, and increased capability for nitric oxide production have all been cited as mechanisms for instantiating long-term memory. All of these mechanisms can easily be seen as sensitive to the amount of time an activity pattern was present.