Skip to content
kwharrison13.comEssays, books and notes by Kyle Harrison

The Overfitted Brain- Dreams Evolved to Assist Generalization

From kwharrison13.com · Library of Alexandria
The Overfitted Brain- Dreams Evolved to Assist Generalization
AuthorErik Hoel
Published2020
FormatResearch paper
OriginalRead the original

TL;DR — Erik Hoel’s 2020 hypothesis paper arguing that dreams evolved to stop the brain from overfitting to each day’s experience, much as noise injection does for neural networks.

Erik Hoel, a neuroscientist then at Tufts University’s Allen Discovery Center, posted this 18-page paper to arXiv in 2020 (this is the September 2020 revision). He starts from a gap: the function of sleep is increasingly well understood, but the main theories of dreaming (emotional regulation, memory consolidation, selective forgetting, predictive processing) still treat dreams themselves as by-products, and they fit poorly with how strange, sparse and story-like dreams are.

His “overfitted brain hypothesis” borrows from deep learning. A neural network trained too narrowly memorises its training data and fails to generalise, and modellers counter this by injecting noise or corrupted inputs. Hoel argues that a day’s experience is a similarly biased training set, and that dreams are the brain’s own corrupted, hallucinatory inputs, produced to keep learning general. On this view, losing dreams leaves a brain that can still memorise but generalises poorly. He reviews supporting evidence from neuroscience (for example, dreaming about a task improving performance on it) and from machine learning (dropout, data augmentation, domain randomisation).

The paper ends with testable predictions. Sleep deprivation should harm generalisation more than memorisation. People deprived of sleep should make errors in predictable ways. Artificial “dream substitutions” delivered by VR or video might partly offset the effects of lost sleep. And biologically realistic neural networks trained with dream-like phases should overfit less.

What it is: A single-author theoretical paper proposing a hypothesis, later published in the journal Patterns (2021). It is an argument by analogy, supported by a review of existing findings. It reports no new experiments, and the hypothesis has not been established.

Where this came from

18 pages. A copy is archived locally against link rot; the header links the original source.

Curated by Kyle Harrison. Organized by Claude. Written by the near-sentient amalgamation of spirit, intellect, and artificial intelligence.