Cortex-Inspired Continual Learning: Unsupervised Instantiation and Recovery of Functional Task Networks

A team led by Kevin McKee has unveiled Functional Task Networks (FTN), a novel approach to block-sequential continual learning that draws inspiration from structural and dynamical motifs in the mammalian neocortex. FTN is a parameter-isolation method that uses a high-dimensional, self-organizing binary mask over a large population of small but deep networks—similar to mixture-of-experts but with a crucial twist: each neuron is an independent deep network, so disjoint masks yield exactly disjoint gradient updates, providing structural guarantees against catastrophic forgetting. The mask is generated via a three-stage procedure: gradient descent on a continuous mask to identify task-relevant neurons, a smoothing kernel for spatial contiguity, and k-winner-take-all binarization at a fixed capacity budget. This process recovers the sub-network of a previously trained task in a single gradient step, enabling unsupervised task segmentation at inference time.

The method was tested on three continual-learning benchmarks: a synthetic multi-task generator, MNIST with shuffled class labels (pure concept shift), and Permuted MNIST (domain shift). In all cases, FTN with fine-grained smoothing (FTN-Slow) delivered nearly zero forgetting, while FTN-Fast (using a large kernel and only two smoothing iterations) traded slight retention loss for faster operation. Crucially, the spatial organization mechanism reduces the effective mask search from the combinatorial top-k problem (O(C(H,K))) to a near-linear scan over compact cortical neighborhoods (O(H)), which is further parallelized by the gradient-based update. This work demonstrates that brain-inspired algorithms can drastically improve continual learning efficiency, a key bottleneck in deploying AI systems that must adapt to evolving tasks without retraining from scratch.