Dreaming in Many Worlds
Examining the topic of dreams, generalization and overfitting.
In Erik Hoel's paper, "The Overfitted Brain: Dreams evolved to assist generalization", Hoel proposes that dreams may have a specific purpose -- for the mind to sample a vast realm of possibilities related to a topic of interest, exploring fringe possibilities as well as good possibilities. This process of exploring the realm of possibilities in our dreams leads to improved "generalization" of our understanding about a topic, or a skill.
Hoel proposes that humans (and AI neural networks) have an issue when they learn new information and new skills -- This issue is called "overfitting". Overfitting is when a person is confronted with a new problem, or wants to master a new skill, they will practice and try to solve the best way to overcome the problem, but the various ways in which they go about solving the problem will be somewhat limited. They will focus in on some successful variants, and some not successful variants. They may fall into an intellectual rut, or develop bad form, due to learning non-helpful patterns. These non-helpful patterns are "overfitted" in our conscious mind, and we get stuck on them.
Learning to Shoot Free Throws - Day One and Night One
Take for instance, a person learning to shoot at basketball from the free throw line. If you're just now learning to shoot the basketball, and spend all day doing so, you will improve over the day, and you will learn (via cognitive and muscle memory), the specific flow and pattern of motions that lead to making baskets. But you'll also fall into some patterns of failure, where some of the pattern is corrupted by poor form or distraction, that leads to missing the basket. At the end of your first full day of practice, you might be a 40% free throw shooter. Your body and mind has learned much during the day, and remembers some good patterns and some bad ones. The bad patterns, that you may repeat because you have learned bad form, could be considered "overfitting" -- Overfitting is a remembered pattern that is not helpful to your progression.
In Hoel's prognosis, if you've spent all day practicing free throw shooting, and go to bed thinking about your technique, your mind will most likely re-run the basketball shots in dreams, and explore the various possibilities for movement and the patterns of muscle and focus. In dreams, you will "re-run the simulation" an indefinite amount of times, and your sub conscious mind will practice and try a variety of different patterns... many of which might produce ridiculous results in dreams and some of which might produce perfect shots. This process of the mind trying out a variety of possibilities leads to "generalization" of the knowledge/memory of how to shoot a basketball. When you wake up, you might focus in on the fact that your dreams were especially bad (dreamed of missing shots), or especially good (dreamed of making shots)... but in both cases, you conscious mind has worked to generalize the muscle memory and cognitive memory, which can help to break any bad form / bad patterns which you might otherwise exhibit going forward. As you head back to the court on your second day, you may experience that after a nights sleep, you quickly improve the next day, and have less of a chance to fall into any bad form / bad patterns that you may have fallen into previously.
Hoel, in his paper, says that this is the reason we dream is to "rerun the simulation" of things we are trying to figure out. By rerunning all the possibilities, even the ridiculous ones, our conscious minds can generalize knowledge of a topic, leading to a better understanding of what is productive and what is not. Bad patterns will be dissolved / diluted by generalization, and good patterns may be improved by variance and juxtaposition.
The same process is true of more intellectual pursuits -- if we're learning a game like chess, or trying to solve some mystery of the universe, if we focus on the topic in dreams, our conscious mind will explore all the possibilities, leading to generalization and dilution of potentially bad patterns, and also provide us with new / novel juxtapositions of good ideas which may lead to improving our good patterns.
The Consciocentric View of Dreams
This analysis by Hoel perfectly fits within the model of Consciocentrism -- Consciousness is constantly traversing this plane of Time * MWI (Many Worlds Interpretation of Quantum Mechanics) When we dream, our conscious minds are allowed to wander far away from the moment of now, and explore the realm of many worlds possibilities and time. Our consciousness runs, and re-runs simulations of possible actions and outcomes and juxtapositions of events in time and many worlds. We store the information in memory, and when we wake up, we have expanded our conscious view of given topics. We have tested both probable and ridiculous extremes, and both types of exploration are valuable, to either generalize our understanding of a topic, by diluting bad possibilities with a variety of extreme examples, or by giving us peeks at new, novel possibilities which we may want to try, in order to improve our skills, or solve our problems.
AI and Deep Learning Neural Networks
Hoel mentions that Neural Networks encounter similar issues to humans when then sample information and try to build knowledge about a given topic. The neural networks encounter some misleading or fringe information that contradicts truth... this is analogous to humans learning a new skill or topic, and experiencing bad form, or digesting information which is untrue, misleading, or not well suited to the problem at hand. Hoel, in his paper, suggests that we may actually want AI and Deep Learning systems to explore the idea of dreams... it may be helpful to task AI and Deep Learning Networks with purposefully exploring various fringe and extreme cases which may fail, or new/novel patterns which may lead to better understanding of what is true. By letting neural networks explore the extremes and the possible "noise" surrounding an issue, it may help to generalize and dilute any noise which would lead to reporting back untruthful information.
Link to Erik Hoel's Paper - The Overfitted Brain: Dreams evolved to assist generalization:
https://arxiv.org/abs/2007.09560