Assessment of Recurrent Spiking Neural Networks on Neuromorphic Accelerators for Naturalistic Texture Classification

This paper presents the implementation of a Recur-rent Spiking Neural Network (RSNN) using surrogate gradient descent for naturalistic textures classification. The implementation choices for the RSNN are limited to hardware-friendly models since it is intended to be integrated into an electronic skin system. Hence, a comparison between the von-Neumman and neuromorphic computing approaches has been assessed in terms of hardware efficiency. The energy consumption per inference of the proposed model is estimated using the KerasSpiking tool built-in NengoDL framework, on three different devices namely: GPU, Intel Loihi, and SpiNNaker. The obtained results indicate that the aforementioned neuromorphic devices achieve several orders of magnitude gains in energy over von-Neumman hardware. Moreover, the proposed RSNN model overcomes similar state-of-the-art solutions in terms of classification accuracy and hardware complexity making it a promising candidate for embedded electronic skin applications.