An RRAM-Based Neuromorphic Accelerator for Speech-Based Emotion Recognition

Abstract

A neuromorphic accelerator for a deep net having RRAM based processing elements has been implemented for emotion detection based on dialect. The proposed accelerator has been trained on the RAVDESS dataset in order to classify different emotion types. The RRAM based swish activation function has been employed to build the neuromorphic accelerator as it consumes less power (476μW), has lower operating voltage (1.23V), and has better performance and output characteristics. The proposed neuromorphic accelerator has been implemented using 1T-1RRAM Processing Elements on ST Microelectronics 28nm FD-SOI, also on Intel i3-8130U CPU and compared with NVIDIA GeForce GPU to highlight the advantages. The proposed accelerator achieves high-performance and consumes less power (1780μW) with on/off rate (13.81) and lower operating voltage (2V). The training accuracy for the FD-SOI implementation is 79.13% and has a learning rate of 0.01 and weight update interval of 1 epoch. This chapter also highlights the importance of the proposed neuromorphic accelerator from Industry 4.0 perspective.