Optimizing reservoir connectivity: A path to high-performance liquid state machines

Abstract

The growing energy demands of artificial intelligence (AI), particularly for deep learning, underscore the need for energy-efficient computational models. The liquid-state machine (LSM), inspired by neuromorphic computing, employs a nonlinear dynamical system and has emerged as a promising candidate. In this study, we investigate the optimal connectivity of the reservoir—characterized by both the number and strength of connections—to enhance LSM performance and energy efficiency. Through systematic analysis, we introduce the spike multiplication factor (λ), a novel metric representing the rate at which spikes propagate through successive neuronal layers at the network level. This factor serves as a key parameter in characterizing LSM behavior. Furthermore, our analysis demonstrates a strong correlation between λ and the reservoir’s chaotic dynamics, as represented by Lyapunov exponents and fractal dimensions. These findings indicate that controlled chaos within the reservoir can significantly enhance LSM performance, offering valuable insights for the design of more efficient neuromorphic systems.