To a very crude approximation, the brain works via Hebbian learning. If neuron A fires immediately before neuron B then the synapse from A to B gets strengthened.

There’s just one problem with Hebbian learning. Carry on long enough and you will eventually end up in a situation where every neuron just fires like crazy all the time. Giulio Tononi’s Synaptic Homeostasis Hypothesis proposes that the function of NREM sleep is to normalise Hebbian learning by weakening synapses.

Some experiments appear to support the synaptic homeostasis hypothesis (e.g.[1]). Other experiments instead find evidence of synaptic potentiation during sleep (e.g. [2]). I’ve seen competing claims that NREM sleep instead normalises Hebbian learning by weakening synapses on neurons that fire a lot and strengthening synapses on neurons that don’t fire very often. Or that neurons use BCM theory rather than Hebbian learning and BCM theory doesn’t need normalising.

I’d like to know what other LessWrong users think about the Synaptic Homeostasis Hypothesis?

[1] Vyazovskiy, V., Cirelli, C., Pfister-Genskow, M. et al. Molecular and electrophysiological evidence for net synaptic potentiation in wake and depression in sleep. Nat Neurosci 11, 200–208 (2008). https://doi.org/10.1038/nn2035

[2] Jaclyn Durkin, BS, Sara J. Aton, PhD, Sleep-Dependent Potentiation in the Visual System Is at Odds with the Synaptic Homeostasis Hypothesis, Sleep, Volume 39, Issue 1, January 2016, Pages 155–159, https://doi.org/10.5665/sleep.5338