Local Pheromone Network Introduces Sparse, Hebbian Learning Without Backprop

A new research prototype from Fu, Chen, and Li challenges the dominance of backpropagation with a biologically inspired alternative: the Local Pheromone Network. The network is sparse — each output unit only reads from a fixed local neighborhood of input units, determined by geometric distance and molecular-tag compatibility. Crucially, it eschews automatic differentiation. Instead, each synapse stores not just a weight, but also a short-term and long-term pheromone trace, plus an optional consolidation state. Training proceeds via a pheromone-weighted Hebbian-style update on a budgeted subset of local synapses, selected based on local error and co-activity. The budget itself adapts online: it shrinks when loss improves and expands toward recently active neighborhoods when loss worsens. Optional mechanisms add structural plasticity, local replay, output masks for partitioned learning, and a target-free local contrastive step.

Early experiments on synthetic tasks demonstrate the network's strengths. It learns local linear rules effectively and can maintain partitioned memories through tags and masks without interference. In conflicting memory tasks, consolidation notably reduces forgetting, while replay helps recover old patterns under continued conflict. The prototype also handles a synthetic long-context hybrid memory task, suggesting potential for sequential or continual learning scenarios. By decoupling parameter updates and using sparse, local learning, the Local Pheromone Network offers a promising path to mitigate catastrophic forgetting — one of the biggest hurdles in lifelong AI systems. While still a small research prototype (10 pages, 4 tables), it opens the door to more robust, biologically plausible network architectures.