Enhancing Liquid State Machine Classification Through Reservoir Separability Optimization Using Swarm Intelligence and Multitask Learning

The Liquid State Machine (LSM) framework addresses supervised learning tasks involving spatio-temporal data streams. It relies on a randomly created, untrained Spiking Recurrent Neural Network (SRNN), called the “liquid,” to map inputs into task-independent representations. A simple readout layer then uses these representations to solve specific tasks. LSM’s computational power arises from two properties: the Separation Property (related to the liquid) and the Approximation Property (related to the readout). This research aims to enhance the liquid’s separation property to improve classification performance and enable multitask learning through swarm intelligence. The study develops a two-phase approach: first, using Particle Swarm Optimization (PSO) to optimize the liquid for distinguishing data streams of different classes in single tasks; and second, extending this optimization to multitask learning with Original Multi-Objective PSO (OMOPSO). Results from experiments on four artificial problems (one of frequency recognition and three of pattern recognition) demonstrate that optimized liquids improve separability and maintain regularized firing behaviors, even with a simple softmax readout layer. On average, the experiments show that our approach outperforms baseline methods across all four artificial datasets when using PSO and achieves superior results on three pattern recognition datasets when employing OMOPSO.