Overfitted AI

There is a very insightful podcast episode - the interview of Andrej Karpathy by Dwarkesh Patel. At 00:53:39 they mention for about 1 second the 2021 paper by Erik Hoel: The overfitted brain: Dreams evolved to assist generalization. It is all just a hypothesis, but I must say I've fallen in love with this hypothesis.

Not much proof exists yet, but Hoel is really on to something. Dreams are all undiscovered yet. Impossible to get into. Brain activity scans of course can do nothing, as it is like trying to figure out the logic of the software by scanning activity of the microprocessor transistors. Impossible.

But surprisingly the recent advancements in AI may get us to better understand how brain works as we try to rebuild an intelligent system in silicon deep neural networks (DNNs). Erik puts forward the overfitted brain hypothesis (OBH):

Notably, one of the most ubiquitous challenges DNNs face is a trade-off between generalization and memorization, wherein as they learn to fit one particular dataset, they can become less generalizable to others. This overfitting is identifiable when performance on the training set begins to differentiate from performance on the testing set.

DNNs memorize a lot. So do our brains during active hours. The challenge then becomes how to convert specific dataset memories into generalized knowledge and experience. Computers have to got there yet. And we hardly have a clue what it takes to get there.

Erik thinks that dreams serve the purpose of going from memorization to generalization. This post-processing elevates the DNN from memorizing the datasets and building some statistical parameters / weights network to a system capable of developing general abstract rules.

The OBH goes even further:

Finally, it is worth taking the idea of dream substitutions seriously enough to consider whether fictions, like novels or films, act as artificial dreams, accomplishing at least some of the same function.

and

The OBH suggests fictions, and perhaps the arts in general, may actually have a deeper underlying cognitive utility in the form of improving generalization and preventing overfitting by acting as artificial dreams.

This really is a beautiful mind (pun intended). To be honest I don't need a proof of the OBH theory. I love dreams, love fiction, love various forms of art. Fundamentally there has always be this feeling of dreams / fiction / art contributing to clearer thoughts.