Restricting plasticity to large, coordinated synaptic updates keeps biological learning cheap — unrestricted backprop would need ~100,000x more synaptic updates in macaque V1

neuroenergeticssynaptic-plasticitygradient-sparsificationbackpropagationlearning-mechanismcross-domain-bridgesparse-updates

Learning is metabolically cheap in the brain not only by accident of biophysics but because the brain actively limits which synapses change. van Rossum & Pache model plasticity-restricting rules — "only modify synapses with large updates" plus confining change to coordinated subnetworks — and estimate that unrestricted backpropagation would require on the order of 100,000 times more synaptic updates than the selective rule in a model of macaque primary visual cortex (V1). Learning energy scales with the number and magnitude of weight changes, so restricting plasticity to the few high-impact synapses is the brain's route to cheap learning — the mechanism behind the low per-event cost that Karbowski measures (claim-synaptic-plasticity-cheap-fraction-of-transmission-energy).

The authors close the loop back to artificial networks: the same restriction rules can be applied to reduce the energy of training artificial neural networks. Framed in machine-learning terms, this is gradient sparsification — updating only the largest-gradient parameters rather than the full weight matrix on every step. It is a biologically motivated argument for the kind of sparse, selective weight updates that AI reaches for to cut training cost, and it stands against the every-weight-every-step character of standard backpropagation (claim-training-inference-compute-asymmetry-mechanism, claim-update-locking-backprop-constraint). It also complements the representational/mechanism-mismatch thesis of backpropagation-gap: the brain that approximates backprop's principles (claim-brain-approximates-backprop-core-principles-ngrad) does so under an energy budget that forbids backprop's literal update volume.