A neural architecture search based framework for liquid state machine design

Liquid state machines (LSMs), also known as the recurrent version of spiking neural networks, have garnered significant research interest owing to their high computational power, biological plausibility, simple structure, and low training complexity. By exploring design space in network architectures parameters, recent works demonstrated great potential for improving accuracy LSM models with However, these are based on manually defined or predefined which may ignore optimization parameters LSMs. In this study, we propose a architecture search-based framework explore parameter automatic dataset-oriented models. To manage exponentially increasing space, adopt three-step search LSMs, including dynamic multiple-liquid multiple layers, variations number neurons each liquid, such percentage connectivity excitatory neuron ratio within liquid. addition, use simulated annealing algorithm implement heuristic search. Two datasets, image dataset NMNIST speech FSDD, were used test effectiveness proposed framework. Simulation results that our can produce optimal The best classification two datasets only 1000 was observed be 92.5% 84%. Meanwhile, connections discovered average, reduced by 56.3% 60.2% separately compared single LSM. Furthermore, total 20% an loss 0.5%.