Fixed-Point Quantization of 3D Convolutional Neural Networks for Energy-Efficient Action Recognition

Abstract

In this paper, 3D convolutional neural networks (CNNs) are simplified to reduce the energy consumption of the action recognition process. Instead of using floating-point weights and input values, which results in a huge amount of processing energy, we introduce a systematic way to quantize all the values of 3D CNNs without degrading the recognition accuracy. Simulation results show that, compared to the baseline CNN architecture, the proposed method significantly reduces the computational complexity as well as the memory requirements.