Reconfigurable Digital Design of a Liquid State Machine for Spatio-Temporal Data

Abstract

Liquid State Machine (LSM) is an adaptive computational model with rich dynamics to process continuous streams of inputs. The generalizable capability of LSM makes it a powerful intelligent engine, with very fast training capabilities. Reconfigurable hardware architectures for spatiotemporal signal processing algorithms like LSMs are energy efficient compared to the traditional Recurrent Neural Netork (RNN) and can also adapt to real-time changes without being application specific. Existing behavioral models of LSM cannot process real time data due to their hardware complexity or fixed design approach. The proposed model focuses on a simple liquid design that exploits spatial locality and is capable of processing real time data. The proposed reservoir hardware is evaluated for epileptic seizure detection with an average accuracy of 85%.