Shimon Ullman

Learning in Deep Networks under Dale's Constraint

This work introduces a biologically motivated neural architecture in which both neural activations and learning signals are represented by non-negative activity, and synapses have fixed sign, while still supporting backpropagation-like learning.