The Overfitted Brain: Dreams evolved to assist generalization (2007.09560v2)

Abstract: Understanding of the evolved biological function of sleep has advanced considerably in the past decade. However, no equivalent understanding of dreams has emerged. Contemporary neuroscientific theories generally view dreams as epiphenomena, and the few proposals for their biological function are contradicted by the phenomenology of dreams themselves. Now, the recent advent of deep neural networks (DNNs) has finally provided the novel conceptual framework within which to understand the evolved function of dreams. Notably, all DNNs face the issue of overfitting as they learn, which is when performance on one data set increases but the network's performance fails to generalize (often measured by the divergence of performance on training vs. testing data sets). This ubiquitous problem in DNNs is often solved by modelers via "noise injections" in the form of noisy or corrupted inputs. The goal of this paper is to argue that the brain faces a similar challenge of overfitting, and that nightly dreams evolved to combat the brain's overfitting during its daily learning. That is, dreams are a biological mechanism for increasing generalizability via the creation of corrupted sensory inputs from stochastic activity across the hierarchy of neural structures. Sleep loss, specifically dream loss, leads to an overfitted brain that can still memorize and learn but fails to generalize appropriately. Herein this "overfitted brain hypothesis" is explicitly developed and then compared and contrasted with existing contemporary neuroscientific theories of dreams. Existing evidence for the hypothesis is surveyed within both neuroscience and deep learning, and a set of testable predictions are put forward that can be pursued both in vivo and in silico.

• The paper proposes that dreams are evolved mechanisms, akin to dropout in deep neural networks, that mitigate overfitting during learning.
• It draws parallels between the sparse, hallucinatory nature of dreams and noise injection techniques used to boost generalization in both brains and DNNs.
• The study integrates neuroscience and deep learning evidence to suggest that REM sleep-driven dreams enhance cognitive flexibility and problem-solving.

The Overfitted Brain: Dreams Evolved to Assist Generalization

Introduction

The paper explores an innovative hypothesis, referred to as the Overfitted Brain Hypothesis (OBH), which postulates that the primary evolutionary function of dreams is to mitigate the brain's tendency toward overfitting during the learning process. This hypothesis is contextualized within an evolutionary framework and takes advantage of recent advancements in understanding both the biological functions of sleep and the operational dynamics of deep neural networks (DNNs).

Understanding Dreams in the Context of Neural Networks

The paper draws a parallel between the challenges faced by DNNs, particularly overfitting, and the learning mechanisms of the human brain. In DNNs, noise injection is a common strategy to enhance generalization capabilities. Similarly, OBH suggests that dreams act as a naturally evolved mechanism to introduce ‘noisy’ or ‘corrupted’ sensory inputs, which can help the brain generalize from daily experiences by preventing excessive memorization of repetitive and biased experiences.

The Phenomenology and Biology of Dreams

Dreams are characterized as sparse, hallucinatory, and narrative, properties that are typically seen as deleterious to straightforward memory consolidation or replay. However, the OBH employs these phenomenological traits to highlight their potential adaptive significance. The sparseness of dreams mimics dropout regularization techniques in DNNs, while their hallucinatory and narrative content serves as a form of domain corruption essential for robust learning.

Evidence from Neuroscience

Compelling evidence for OBH emerges from neuroscience where REM sleep and its facilitation of dreaming are linked to improvement in cognitive performance and problem-solving. Behavioral studies indicate that cognitive tasks benefit more from sleep than perceptual tasks, which suggests that dream-induced generalization has specific impacts on higher-order cognitive activities. Furthermore, dreams about novel but over-learned tasks (like games) align with the conditions under which OBH predicts that dreams are most necessary.

Evidence from Deep Learning

The paper emphasizes parallels in deep learning research, where strategies to counteract overfitting, such as dropout and domain randomization, share functional similarities with dreaming. In DNNs, techniques like dropout (which introduces randomness in inputs) and domain randomization (which alters input distributions) have consistently shown to improve model robustness and generalization, resonating with the proposed role of dreams.

Predictions and Future Directions

OBH generates testable predictions that can be pursued in laboratory settings. Experimental validation could involve tracking the impact of dream manipulation on cognitive and perceptual tasks, as well as modeling dream-like conditions in computational simulations. For practical applications, exploring the effects of artificially induced dream-like states could serve to alleviate the cognitive detriments of sleep deprivation.

Conclusion

The Overfitted Brain Hypothesis provides a cohesive framework linking dream phenomenology with neural network learning strategies, suggesting that dreams are a biological adaptation for enhancing cognitive flexibility through improved generalization. While the hypothesis remains speculative and requires substantial empirical validation, it opens promising avenues for interdisciplinary research bridging neuroscience and artificial intelligence. The understanding of dreams as more than epiphenomena but as crucial components of cognitive optimization processes expands theoretical insights into both biological and artificial systems of learning.