### : Dreams evolved to assist generalisation ### **Summary of the Paper: The "Overfitted Brain Hypothesis"** This paper proposes a novel theory about the biological function of dreams, suggesting that they help prevent the brain from **overfitting**—a phenomenon where a learning system performs well on trained data but poorly on new, unseen data. #### **Key Points:** 1. **Dreams as a Biological Noise Injection Mechanism** - Deep neural networks (DNNs) combat overfitting by introducing noise (e.g., corrupted inputs) to improve generalization. - Similarly, the brain may use dreams (random, distorted sensory experiences during [[Sleep]]) to "regularize" learning, preventing overfitting in neural networks. 2. **The Overfitted Brain Hypothesis** - Dreams arise from stochastic (random) neural activity across brain structures, creating bizarre, novel sensory inputs. - This process helps the brain generalize better rather than just memorize daily experiences. - Lack of dreaming (e.g., due to sleep deprivation) could lead to an overfitted brain—one that learns but fails to adapt flexibly to new situations. 3. **Contrast with Other Dream Theories** - Unlike theories that see dreams as meaningless byproducts (epiphenomena) or emotional regulators, this hypothesis positions dreams as a **functional adaptation** to enhance learning efficiency. 4. **Supporting Evidence & Future Research** - [[Neuroscience]] and AI research both show that noise improves generalization. - The paper suggests testable predictions, including experiments in both biological brains and artificial neural networks. ### **Conclusion** The "overfitted brain hypothesis" frames dreams as a natural mechanism to optimize learning, much like noise injections in AI. This theory bridges neuroscience and machine learning, offering a fresh perspective on why we dream. https://www.sciencedirect.com/science/article/pii/S2666389921000647 https://arxiv.org/pdf/2007.09560#:~:text=Put%20another%20way%3A%20the%20surest,generalize%20performance%20on%20the%20task. `Concepts:` `Knowledge Base:`