Medusa: an event-driven spiking network that learns from reward alone, proposes and validates with two hemispheres, and transfers concepts across tasks

We present Medusa, an event-driven spiking neural network that learns without backpropagation and, at the decision stage, without labels: it only receives "correct / wrong" (dopamine and serotonin signals gating three-factor STDP). A single design - STDP feature detectors, a Go/NoGo habit, and two hemispheres that propose and validate, linked by a corpus callosum - is applied without redesign to handwritten and spoken digits, chest X-rays, chess, checkers and Tetris. We report three main results. (1) By copying two features of biological vision - parallel pathways of different granularity and a graded population code between layers - Medusa reaches 95.2% on MNIST from reward alone, matching Diehl and Cook (95.0%) with about 30 times fewer visual neurons (about 220 versus 6,400); with a single eye, the two hemispheres add 2.0 to 4.5 points. (2) Medusa transfers concepts between tasks that share the meaning of the decision: it plays checkers with what it learned in chess without any practice (45% versus 6% for a blank network), and it recognizes pneumonia in children without having seen a single case, after learning pleural effusion in adults (AUC 0.81 versus 0.53). (3) When tasks share only appearance (shapes -> digits), nothing transfers and inherited knowledge can hurt. Controls with shuffled weights, random pairings and crossed labels show that what travels is meaning. Finally, guided only by dopamine, Medusa discovers by itself which chess knowledge is useful in checkers.