on-off-representations

This paper introduces a biologically motivated neural architecture that respects Dale's constraint—neurons and synapses have fixed sign—while still supporting backpropagation-like learning using non-negative error signals and local Hebbian updates, showing gains on Tiny ImageNet.