The overfitted brain: Dreams evolved to assist generalization

Area of Science:

• Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• The biological function of sleep is increasingly understood, but the purpose of dreams remains elusive.
• Current neuroscientific theories often consider dreams epiphenomena, lacking empirical support from dream phenomenology.
• Deep neural networks (DNNs) offer a new paradigm for understanding complex biological processes.

Purpose of the Study:

• To propose the

Main Methods:

• Drawing parallels between DNN overfitting and potential brain overfitting mechanisms.
• Utilizing concepts from deep learning, specifically noise injection, to explain dream function.
• Reviewing existing neuroscience and deep learning literature to support the hypothesis.

Main Results:

• The paper introduces the "overfitted brain hypothesis" positing dreams as a mechanism to prevent overfitting.
• Dreams are proposed to enhance generalizability by generating corrupted sensory inputs.
• Sleep deprivation, particularly dream loss, results in an overfitted brain with impaired generalization.

Conclusions:

• Dreams evolved as a biological solution to prevent the brain from overfitting during daily learning.
• The "overfitted brain hypothesis" provides a novel framework for understanding dream function.
• The hypothesis is supported by evidence from neuroscience and deep learning, with testable predictions proposed.