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

Spiking reservoir computing is a robust framework for neuromorphic and biological applications, relying on a fixed random reservoir driving a trained readout. To overcome limitations in reservoir initialization, researchers are exploring adaptive, unsupervised plasticity rules to enhance performance. However, identifying which rules work best and why remains a challenge due to the time-consuming and resource-intensive search process.

In this study, researchers utilized a plastic spiking reservoir network to train a version of the Atari game Pong. By systematically characterizing a large family of local plasticity rules previously meta-learned, they discovered that high-scoring rules exhibit distinct patterns. These patterns include strong differentiation between neurons encoding the ball trajectory and background, stable weight dynamics, and consistent readout alignment across time.

These features are not specific to the Pong task and can be applied to other tasks like a delayed-match recognition task.

Furthermore, the performance of these rules can be estimated solely from their parameters. Leveraging simulation-based inference (SBI) and conditioning it on high scores enables researchers to directly sample from promising regions of the rule space. This approach has led to the discovery of rules that surpass the performance of those found in the original distribution, showcasing stable and high-performing configurations.

These findings provide competitive results to classical reservoir computing systems, while simultaneously offering a transparent and interpretable explanation of what makes a plasticity rule effective for neuromorphic hardware and biological computing.