Brain algorithms protect models of reality

From supermemo.guru
Jump to navigation Jump to search

This article by Dr Piotr Wozniak is part of SuperMemo Guru series on memory, learning, creativity, and problem solving.

Modelling reality

Human brain has an immense power to spawn and protect coherent models of reality. All our behavior is determined by this abstract generalization rather than by the actual reality. At any given moment, our overall model of reality is confronted with sensory input data, and the map of activations in the brain. The activations and sensory data are integrated using the model of reality to determine the next decision or the next action.

To ensure efficient function of the brain and the body, the model of reality should be coherent and consistent. It is also helpful if it correctly reflects the reality. Incorrect modeling leads to errors, incl. death. A child may believe it can outrun a train. This may provoke risky behaviors. This may also lead to a prompt correction of the error in the model.

Two layers of model protection

We have evolved a number of imprinted algorithms for protecting models stored in memory. Those mechanisms work at the neural level (e.g. confirmation bias), and at the social level (e.g. resistance in coercive learning). A common myth says that those protections are based on faulty algorithms. Cognitive biases, incl. the confirmation bias, are considered errors of the mind. In reality, protection of models leads to diversity of models in a population, which leads to a competition between models, and the evolution of models. The crystallization of collective knowledge undergoes processes that are very similar to those that happen in individual brains (competition, interference, forgetting, revaluation, crystallization, stabilization, etc.). The entire evolutionary process underlies human collective intelligence, and the progress of mankind.

The two essential layers of model protection ensure model stability:

  • brain level: generalization is an inherent property of neural networks. We build models that tend to reject inconsistent information. This result is a healthy phenomenon known as the confirmation bias
  • social level: social interaction leads to group polarization. This means it tends to stabilize in one of the two extreme states: either (1) the meeting of minds (to confirm models) or (2) intellectual combat (which favors collecting evidence to confirm models)

Generalization can be faulty, but it improves fast thinking and makes further generalizations easy. Bistable interaction between models at the social level, leads to tribalism (formation of clans and coalitions), accelerates generalization, favors coherent learning, and contributes to a clash of models that may, in extreme cases, lead to revaluation. In matters of high weight, on rare occasions, people change their minds under the pressure of the overwhelming evidence to the contrary.

An individual point of view receives a natural protection at the neural and at the social level

Pain of broken models

Breaking up models is unpleasurable. If the model is strong, and its valuation is high, the pain may be significant. This results in an "emotional attachment" to personal beliefs. This seemingly irrational phenomenon is part of the model protection scheme.

Breaking up a faulty model will often lead to a rich compensation when a new model can be established. This is why it is much easier to transition between models than to break up an existing model with nothing new in its place. The breakup of a strong model may lead to a painful realization: "I have been wrong all along" or "I have wasted my youth/life on a wild goose chase". This is also a reason why a wrong model based on limited data is better than no model. A wrong model can serve as a skeleton of generalization upon which inconsistencies (errors) can be discovered. When a model is missing, coherent learning is inhibited, and the progress stalls. If there is a great deal of data that supports several alternative models, keeping all models in memory may be unavoidable, however, that approach is costly and error-prone. An early choice in favor of one of the models may accelerate progress (incl. falsification of the choice).

A wrong model is better than none

Smart people are stubborn

The smarter the individual, the stronger the algorithmic processes of generalization, and the stronger the socially-driven protection mechanisms. It is much easier to persuade an individual whose knowledge is less extensive, less stable, and less crystallized. Adult populations around the world may seem highly persuadable. The factors that weaken the mind are numerous: separation anxiety in daycare, limits on freedom at school, learned helplessness, depression, rat race, bad health, etc. The two key forces that lead to the weakening are the suppressed learn drive (less learning), and incoherent learning. Without the continuous flow of new knowledge, memory structures wither due to forgetting and interference. Instead of lifelong learning, we may experience the loss of the joy of living.

In the name of healthier and happier populations, we need to protect the healthy force of the learn drive, lifelong free learning, and develop tolerance to diversity of opinion. Debating points of view is healthy, however, incessant critical bombardment of weaker models may lead to their break up, without a chance to put in new structures that could underlie the recovery. Tolerance facilitates coexistence. The clash of models is welcome. However, the optimum scenarios in a clash of models are no different than solo learning: it is always recommended to tackle one issue at a time. It is better to conclude one debate before commencing another.

We need to celebrate strong models (points of view), even if they differ from ours. When someone tells you "you are the most stubborn person in the world! You are impervious to argument", take it as a potential sign of your own high intelligence. Models do not need to be correct (see: Value of wrong models). The myth of "the only correct model" we learn at school where only "one truth" is acceptable. That "one truth" often appears to be just a point of view, or a specific interpretation, e.g. of a historic fact. Diversity of models needs to be protected and cherished. Unfortunately, we do not seem to have a brain algorithm for protecting diversity. Tolerance needs to be acquired by learning in the same way as skepticism (i.e. immunity to fake news). Possibly, it is the modern connected world that necessitates that one extra layer of protection for models. We have evolved in conditions of lesser social connectivity. Diversity is precious, but in a connected world, it can turn out overwhelming.

With a bit of self-discipline and training, we can develop a degree of tolerance for diversity. This may turn out helpful to methodically resolve contradictions between diverse models: one at a time. My favorite model to clash with is the model of "good school".

We need to master the tolerance of diversity. In the connected world, we might be missing a natural mechanism for the protection of diverse models

I apologize for my incorrigibility

If you happen to read some of my text, you may have an impression that I am impervious to argument too. I am proud of it. When I get blasted, I double down. Each time I hear of my stubborn stance, I look for good examples in which I was easily convinced with a coherent argument. Invariably, I hear back "That's not a good example. This is obvious". In other words, I do not reject obviously valid claims. I reject those that clash with my models, and I am happy my models are strong, even if they may occasionally turn out wrong. If you see my error, let me know