PORTFOLIO

aThe study

Question

Does the speed of firing-rate homeostasis decide whether a spiking network can keep learning new tasks?

Setup

Two-layer spiking networks trained on 150 permuted-MNIST tasks in a row, with per-neuron homeostasis at different speeds.

Result

Homeostasis on the timescale of about one task kept the network trainable late in the stream. Fast homeostasis and none both lost that ability.

Evidence

Slow beat both in 5 of 5 seeds. New-task learning (AUC) was 0.71 against 0.45 and 0.43.

bLate-stream accuracy

• 0.88slow homeostasis
• 0.77fast homeostasis
• 0.79no homeostasis

aThe study

Question

How close can local learning rules get to backpropagation in streaming reinforcement learning?

Setup

A streaming actor-critic agent on MuJoCo tasks, batch size 1 and no replay, with hidden layers trained by local rules.

Result

Fixed feedback weights that share only the signs of the forward weights came within about 10% of backpropagation.

Prediction

Sheppard's arcsine law from 1899 predicted how often local and true updates agree in sign: 0.692 predicted, 0.695 measured on Humanoid.

bFinal return, as a share of backpropagation

Pooled over four MuJoCo tasks. The interval for sign-sharing feedback is 0.87 to 1.03.