Progressive networks & parameter-expansion continual learning (Phase 177.6), in PyTorch, CPU-runnable.
Components#
- Progressive Networks (
progressive.py, 177.6.1.1) — Rusu & Hadsell 2016. A growing stack of columns; adding a column freezes all prior columns, and each column reads lateral adapters from prior columns' previous-layer activations. The backward pass only touches the new column, so earlier tasks are preserved bit-for-bit (verified: task-0 predictions are identical after training column 1). - Column-pruning distillation (
column_pruning.py, 177.6.1.3) — distil rarely-used frozen columns into a compact shared trunk with per-task heads (logit KD + label CE), gated on measured skill retention. - Dynamically Expandable Networks (
den.py, 177.6.2.1) — Yoon et al. 2018. Selective retraining with L1, then loss-triggered width expansion training only the new units, old units frozen. - PackNet (
packnet.py, 177.6.2.2) — Mallya & Lazebnik 2018. One backbone, iterative magnitude pruning, per-task weight ownership + per-task biases, and the two-phase train→prune→re-train schedule. Earlier tasks preserved exactly. - Supermasks (
supermask.py, 177.6.2.3) — Zhou et al. / edge-popup. Frozen random weights, per-task learnable scores, top-k binary masks with a straight-through estimator; different tasks route through different masks with zero weight mutation.
Tests#
python3 -m pytest tests/ -q — 15 tests: progressive no-forgetting + frozen
priors + lateral presence, distillation retention + compression, DEN expansion
on hard tasks / no growth when easy, PackNet exact preservation + disjoint masks
- both-tasks-learn, and edge-popup fitting frozen weights + distinct per-task masks + STE gradient routing.
Not locally actionable#
177.6.1.2 (integrate progressive columns with the Galatea VLA, Phase 176.7.1) requires that VLA policy and is left unchecked; the column mechanism it would use is complete here.