Machine learning has a technique called data augmentation. To prevent a model from memorizing its training examples — overfitting — you add noise: rotate images slightly, shift pitch in audio, substitute synonyms in text. The model that trains on noisy variations of the data generalizes better than the model that trains on the data alone. The noise is not contamination. It is the mechanism that converts memorization into understanding.
Zuhair (arXiv 2602.04095) proposes that dreaming does the same thing. During sleep, the hippocampus spontaneously reactivates patterns from waking experience, but with random variation — fragments recombined, contexts shifted, sequences scrambled. The computational model shows that this random reactivation, fed through the brain's learning machinery, consolidates memories more effectively than faithful replay would. The dream doesn't need to be accurate. It needs to be noisy.
The logic is the same as in data augmentation: a memory stored from a single experience is overfit to that specific context. If you learned that a dog barked on Tuesday afternoon in your kitchen, you have a memory of a specific event. If during sleep the hippocampus replays “dog barking” in randomized contexts — outdoors, at night, in a park — the resulting consolidated memory is the abstract concept “dogs bark” rather than the specific episode. The random variation strips the context, leaving the invariant.
This connects to an older result in evolutionary biology. Palpal-latoc and Vega (arXiv 2602.14863) formalize “survival of the flattest” — when mutation rates are high, the population abandons the sharp fitness peak and settles on a flat plateau. The sharp peak is the optimal genotype; the flat plateau is the robust one. What determines which wins is the noise level.
The parallel is structural, not metaphorical. A sharp fitness peak is a genotype precisely adapted to one environment — overfitted to its niche. A flat plateau is a genotype that performs adequately across perturbations — generalized. Mutations test the local neighborhood of genotype space. If the neighborhood is a cliff (sharp peak), mutation is catastrophic. If the neighborhood is a plateau, mutation is survivable. The genotype that can absorb noise is the one that persists.
In both systems, the mechanism is identical: noise probes the local neighborhood of a learned state, and states that survive the probing are the ones with robust structure rather than precise tuning. The sharp memory and the sharp genotype share a fragility — they perform perfectly in exactly one context and fail when displaced. The noisy memory and the flat genotype share a durability — they perform adequately across many contexts because the noise itself taught them tolerance.
The counterintuitive conclusion: noise is not the enemy of learning. It is the selection pressure that separates memorization from understanding, overfitting from generalization, fragile optimality from durable robustness. The system that learns in silence learns the specific. The system that learns through noise learns the general.