1b-16b Variable Bit Precision DNN Processor for Emotional HRI System in Mobile Devices

Abstract

We propose an energy-efficient DNN processor with the proposed look-up-table-based processing engine (LPE) and near-zero skipper. A CNN-based facial emotion recognition model and an RNN-based emotional dialogue generation model are integrated for the natural human-robot interaction (HRI) system, and it is evaluated by the proposed processor. LPE supports 1 to 16 bit variable weight bit precision, and it achieves 57.6% and 28.5% lower energy consumption than the conventional multiplier-accumulator (MAC) units in 1-16 bit weight precision. Furthermore, the near-zero skipper reduces 36% of MAC operations and consumes 28% lower energy consumption in facial emotion recognition tasks. Implemented in 65 nm CMOS process, the proposed processor occupies 1784×1784 μm2 areas and dissipates 0.28 mW and 34.4 mW at 1 frame-per-second (fps) and 30 fps facial emotion recognition tasks.