Memory formation theory

What is the basic structure of our brains? How are our memory formed? Can we influence that? Do we need to rely on mnemonic devices each time? Will our memories get lost in translation? Today we will try to examine the most basic layer of memory. This article is important, so I will try to be brief and simplify everything. Check out here, here, and here.

What is memory?

Our brain has a lot of neurons. 100 billion neurons is a good estimate. This is roughly for times more than transistors in GeForce RTX 3080 GPU. Only each neuron is significantly more complex than a single transistor, and there are significantly more connections between the neurons than in a GPU. Transistors are faster by a factor of 10mil, so AI intelligence is getting close to ours.

These neurons are connected to each other via slow connections, but many of them. Not all possible connections are active. Moreover, there is some information stored within the neurons themselves. In a way, our brain is the ultimate mindmap.

As a signal is passed through the brain, it goes through the most active connections being modified in each neuron. In a way, this is similar to our convolutional neural networks.  The strength of each connection and the operation applied by each neuron change very slowly. This is the basic mechanism of memory.

How do the associations work?

We have very little control over 80% of our brain. Automatic processing is faster and better than deliberate thinking. Roughly 20% of our brain neurons are dedicated to deliberate thinking, and it is our neocortex. The neocortex takes 76% of the entire brain volume.  So it is structurally different. The neocortex is built like a processing chip. It is basically a two-dimensional sheet with roughly six different layers of neurons.

This neocortex “sheet” is folded many times to get inside our scalp. Between different areas of the “sheet” run very fast neural connections covered in white material. These connections form our associations. Not all associative connections are used. When a connection is used more often, it gets more white cover, becomes faster and stronger.

I deliberately try to avoid medical terminology and be as intuitive as I possibly can. It matters for example if you want to understand why associations work better between interconnected subjects or why each language comes with its own set of associations.

Gender differences

Women have a smaller neocortex than men because they are generally smaller. There are roughly 16% fewer neurons in the female neocortex. So women are less effective in visual processing and math. Women are however better in languages and emotions, which require fewer neurons for meaningful analysis.  Neanderthals had brains 10% larger than modern humans, so bigger is not always better. Some studies show that transsexuals actually have the brain qualities of the sex they feel they belong to.

There is no proof that one gender remembers better than the other. Some studies show that musicians have a better working memory. Most memory champions are men. The fastest reader alive is a woman, but several fast readers after her are men.

With age, our brains shrink and the male brain becomes more similar to the female. We see increased linguistic activity, but at the same time, visual processing and math become less effective. These differences, often 10% to 20%  are negligible when compared with the effects of training and strategy.

Encoding

When we remember something this means several things. First of all, we need to notice the information we want to remember. This requires a certain amount of focus.  Typically we will filter out everything that is not special and does not stand out. The filtering step reduces the amount of information we need to process, but occasionally we through out something we need. So make it special to remember it.

Then we need to encode the information we left with. Usually, we remove some specific details, again to remember less information. If we remove too many details, we are left with nothing to remember. So some details must stay, and these details must be very specific and characteristic. Like the red dress on a girl in Schindler’s list.

If you took one of our courses, you will probably use some sort of dual coding: define what you want to remember as three words and an image. This is very effective, but not very natural. Naturally, we simply encode the new information into the relevant neuron and its connections. We do not really know how much information a neuron can hold, but it is pretty massive. As we rarely learn disconnected information, the new information we encode usually updates closely connected clusters of neurons.

Retrieval

Already during the encoding stage, we are aware of future retrieval. We ask ourselves: what would remind us of the stuff we just learned? Quite often we quiz ourselves in a sort of forced retrieval. This strategy is intuitive. Memory masters often do it more than the rest of us, because it is very important.

When we retrieve information, we work very much like online search engines and try to find the match closely associated with the pattern we use for the search. The pattern can be verbal, or it can be an image, or something more complex encoded by our brain. Since the encoding mechanism of the brain changes slowly, a close match is usually found if the pattern is specific. A generic pattern generates too many similar and confusing matches. So we usually look for something very specific. Just like we do in our web search. We typically use three words for a good query.

Going back to the encoding stage, we add this very specific pattern that will facilitate later retrieval. Anna suggests asking questions. I add easily recognizable details or logical symbols.

Context

We remember things in a certain context. Context works like a namespace in programming: additional information that is automatically added to all of our retreivals. We do not need to add it specifically each time, reducing the complexity of retrieval and improving its speed.  If we try to retrieve information outside of its context we may get a mismatch. This is like searching for information on the wrong website.

Context is usually hierarchical. We have some more abstract idea or field of knowledge and within it more specific stuff. This ontology works like our address. We can search for Lev Goldentouch in Israel and find me, but if you know the city and the street, you will find me faster and there will be only one person with that name.

Probably, context is the most underutilized of mnemonic devices. It does not help in memory championships, but it is super important in actual learning. For example, when we go from one language to another even our associations change.

Reinforcement

A single exposure to any sort of information is not sufficient for learning.

When we are exposed to rare and dramatic events we usually replay them in our mind. This can be traumatic and uncomfortable, but it is a very effective learning mechanism. We do something similar when we train AI and repeat rare inputs many times.

If the events are not dramatic, this mechanism does not kick in. We actually need to be exposed to the relevant information many times. The easiest way to achieve this while doing research, is reading a large number of closely related documents. A more effective way includes reading significantly fewer documents but writing some. Spaced repetition is probably an inferior way to achieve this as we do not gain additional insights per repetition, but it is straightforward and it works.

Automatic memories

While we can do quite a lot with our more deliberate memorization, most of our training is invested in forming automatic memories. We do not forget how to walk, how to breathe, or how to ride a bike.  Most of our neurons are not in the neocortex. Our automatic processing is significantly faster and more effective than deliberate processing.

The hands-on training is just that. Doing the same things over and over, until it gets effortless. The most important part here is not learning something with mistakes, because refinement and relearning may be harder than the original learning. We have something like that in AI, where the learning rate drops as we learn more. So make sure to learn the fundamental skills well.

Practice slowly in the beginning to remove any mistakes. Overlearn them. Make them automatic. Hire a coach to oversee the process…