Learning in Deep Networks under Dale’s Constraint

Name: Learning in Deep Networks under Dale’s Constraint
Uploaded: 2026-08-10T07:08:13.003361+00:00
Duration: 1072 s
Description: Tom, Jane, Lu, Meng, and Lalam discuss "Learning in Deep Networks under Dale's Constraint" by Abel and Ullman. The paper introduces an ON/OFF channel motif that lets Dale-compliant neurons encode signed values, and proves a local Hebbian rule recovers exact backprop updates. On Tiny ImageNet, the constrained model beats vanilla convnets.

Tom, Jane, Lu, Meng, and Lalam discuss "Learning in Deep Networks under Dale's Constraint" by Abel and Ullman. The paper introduces an ON/OFF channel motif that lets Dale-compliant neurons encode signed values, and proves a local Hebbian rule recovers exact backprop updates. On Tiny ImageNet, the constrained model beats vanilla convnets.

Open the full episode