General-purpose LSM learning processor architecture and theoretically guided design space exploration

This paper presents a general-purpose liquid state machine based neuromorphic learning processor with integrated training and recognition for real world pattern recognition problems. The proposed architecture consists of a generic preprocessor and one or multiple task processors. The pre-processor, or the reservoir, consists of a recurrent spiking neural network with fixed synaptic weights. Task processors are light weight and comprise a set of readout spiking neurons with plastic weights, which are tuned by a biologically plausible supervised learning rule. Importantly, we leverage the unique computational structure of the reservoir for highly efficient implementation of multiple tasks on the same learning processor. A novel theoretical measure of computational power, which is strongly correlated with the true learning performance, is proposed to facilitate fast design space exploration of the recurrent reservoir. We demonstrate the application of our processor architecture by mapping four recognition tasks onto a reconfigurable FPGA processor platform.