HebbGate: Local Reward‑Modulated Gating for Continual Learning

Andreas Papachristodoulou, Christos Kyrkou, Stelios Timotheou, Theocharis Theocharides

Keywords: Continual Learning, Incremental Learning, Hebbian Plasticity, Local Learning Rules

TL;DR: HebbGate applies a local, reward‑modulated Hebbian rule to create sparse task‑specific subnetworks, delivering fast continual learning with minimal extra memory.

Abstract:

Neural networks that learn continually should acquire new tasks without revisiting old data and with small per-task overhead. In parameter-isolation CL, existing approaches typically learn dense task masks via backpropagation, which couples mask learning to the backbone optimiser, adds training compute, and inflates memory with extra mask parameters.
We introduce HebbGate, a parameter-isolation method for continual learning that uses local, reward-modulated gates in place of backpropagated masks. Crucially, each task adds just one scalar per channel (not per weight), keeping memory growth tiny and the masks interpretable. A utilisation penalty discourages reuse of over-popular channels, and a $\kappa$-decay capacity warm-up lets new tasks explore larger masks before annealing to the target sparsity, mitigating order bias and improving forward transfer.
On CIFAR-100, Tiny-ImageNet-200, and ImageNet-100 with ResNet-18, HebbGate achieves best-known exemplar-free Class-IL final accuracy $A_{\text{last}}$ while a variant with task-specific BatchNorm further improves both $A_{\text{last}}$ and incremental accuracy $A_{\text{inc}}$ at the cost of only two additional scalars per channel per task. Additional experiments on Permuted-MNIST, Split-CIFAR-10, and lower-capacity backbones confirm that HebbGate’s gains extend beyond a single architecture or dataset. Overall, HebbGate offers a lightweight, transparent alternative for exemplar-free, single-head continual learning.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 2140