Local learning rules that scale to deep networks
A. Sato · Neuro Group
arXiv:2604.02218
Backpropagation requires a global backward pass no cortex could run. We study local learning rules, updates computed from information available at each synapse, and ask how far they scale.
On standard benchmarks, our rules train deep networks to accuracy approaching backpropagation, while offering a path to hardware that learns online and cheaply.
This is a plain-language summary. The full manuscript and supplementary material are available on request.