Evaluating Readout Strategies in Liquid State Machines: Coding Schemes, Classifiers, and Separability Metrics

Abstract

Translate Article

Liquid State Machines (LSMs), a form of Reservoir Computing (RC) based on spiking neurons, have shown promising results for tasks involving complex, temporally structured data. However, research on the impact of coding schemes, classifier choice, and separability metrics applied to LSM remains insufficient. In this work, six coding schemes, six classifiers, and three separability metrics across two synthetic benchmark tasks (Frequency and Pattern Recognition) are systematically evaluated. This allows us to contrast readout strategies and assess how well separability metrics align with classification accuracy. Our results showed that rate-based coders combined with linear classifiers yield the most accurate performance. Among separability metrics, K-Nearest Neighbor Overlap Score (KNN-OS) provides the most consistent correlation with accuracy, while Silhouette Score (SS) and Davies-Bouldin Index (DBI) display important limitations. These findings contribute to a clearer understanding of readout design in LSMs and provide practical guidance for selecting coding strategies, classifiers, and evaluating metrics in future applications.