Sleep as Mandatory Pruning of the Brain's Generative Models

The Free energy principle offers a reframing of sleep that is anything but trivial. The variational Free Energy Bound decomposes into two terms: accuracy and complexity. Accuracy measures how well the brain's generative model fits incoming sensory data; complexity measures how far that model has drifted from simpler prior beliefs — essentially, how over-parameterized it has become. During waking life, both terms compete: the brain must stay accurate to the sensorium while also keeping complexity in check. But during sleep, sensory and motor channels are largely shut down. The accuracy term drops out. What remains is pure complexity cost.

This means sleep provides a privileged window for complexity minimization — pruning synapses that encode redundant or overly specific associations, simplifying generative models, and consolidating only the Causal Structure that genuinely supports generalization. The analogy to machine learning is direct and illuminating: an over-parameterized model that fits training noise perfectly will fail catastrophically on new data. The brain faces the identical Overfitting problem, and sleep is its regularization mechanism. Dreaming, from this perspective, is the offline rehearsal of the model's essential causal architecture — a process of distillation rather than mere replay.

What makes this account especially compelling is its convergence with Giulio Tononi's Synaptic Homeostasis Hypothesis, which arrives at a structurally similar conclusion — that sleep exists to renormalize synaptic weights accumulated during waking — from entirely independent neurobiological reasoning. When two frameworks with different starting assumptions and methodologies converge on the same functional story, it constitutes strong abductive evidence that the underlying mechanism is real.