Learning to play with spikes: characterizing, predicting, and engineering unsupervised plasticity rules for spiking reservoir computing

High-scoring local plasticity rules in a plastic spiking reservoir network for Atari Pong are linked to strong neuron differentiation for ball trajectory versus background, with stable weight dynamics and consistent readout alignment.

• Why it matters: Reservoir computing performance hinges on initialization, and adaptive unsupervised plasticity could overcome this, but it has been unclear which rules work, why, and how to find task-suited rules efficiently.
• What they did: The study meta-learns and systematically characterizes a large family of local plasticity rules, then tests whether rule performance is predictable from rule parameters and whether mechanistic signatures transfer to delayed-match recognition.
• The result: Conditioning simulation-based inference on high Pong scores enables sampling from promising rule-space regions, yielding rules that exceed prior-distribution performance and provide interpretable, hardware-relevant configurations.

AI summary of the abstract · may contain errors · not medical advice · read the paper.