Back to Research Labs
modular-rl
Continual Learning Plasticity Demo

Permuted MNIST: Loss of Plasticity vs. SWR & CBP

Watch standard deep neural networks (Vanilla SGD) suffer from Loss of Plasticity across 50 sequential task permutations, while Selective Weight Reinitialization (SWR) and Continual Backprop (CBP) maintain ~95% accuracy indefinitely.

Plasticity Simulation Controls (Task 1 of 50)

Adjust SWR reinit threshold k and CBP replacement rate ρ.

SWR Threshold Factor (k):1.0e-5
CBP Replacement Rate (ρ):1.0e-4
Task Step Speed:80 ms
Vanilla SGD (Base Model)
95.9%

Suffers from severe Loss of Plasticity as dead units accumulate and features collapse.

SWR (Selective Weight Reinit)
95.5%

Periodically reinitializes low-utility weights and resets momentum, preserving network health.

CBP (Continual Backprop)
94.8%

Continuously replaces mature low-utility neurons and re-routes bias projections.

Permuted MNIST: Online Accuracy (%) Over 50 Sequential Tasks

Online evaluation accuracy before training on each 784-dim input pixel permutation task.

Vanilla SGD
SWR
CBP
100%80%60%40%20%