Functional Task Networks: Cortex-Inspired Spatial Parameter Isolation for Continual Learning

Continual adaptation is hardest under concept shift, where successive tasks share an input distribution but require incompatible outputs. We introduce Functional Task Networks (FTN), a parameter-isolation method that routes each task through a sparse subset of independent components without requiring a task ID at inference. Given a small labeled support batch, FTN constructs a binary mask by optimizing mask logits, applying spatial smoothing on a component grid, and selecting a fixed-capacity set of active components. The same cold-start procedure is used to allocate components during training and to recover them at evaluation, so task inference is handled by the adaptation mechanism itself. Disjoint masks yield disjoint gradient paths, giving a structural no-forgetting guarantee under standard optimizer precautions. We instantiate the same idea in LoRA by replacing a single low-rank adapter with a mask-routed pool of rank-r components. On synthetic concept-shift benchmarks, FTN substantially reduces forgetting relative to shared-parameter baselines while preserving task performance. On a six-task distilgpt2 style-transformation stream, LoRA-FTN reduces forgetting from +7.37 to +0.07 nats under support-batch mask recovery, approaching an oracle disjoint-adapter reference without using task labels.