Lightweight End-to-End Stress Recognition using Binarized CNN-LSTM Models

Abstract

In this paper, we propose a novel end-to-end stress recognition model by combining binarized convolutional neural network (CNN) and long short-term memory (LSTM) models. Based on the previous CNN-LSTM model using electrocardiogram (ECG) and respiration (RESP) signals, we newly apply the bandit-based hyperparameter optimization to find more accurate solutions. Analyzing the computational costs of the accuracy-aware model, we also introduce advanced memory-reduction techniques with downscaling and binarization for realizing the cost-efficient stress recognition solution. As a result, compared to the state-of-the-art methods, the proposed model reduces the memory size, the inference latency, and the energy consumption by 93 %, 39 %, and 42 %, respectively, while even increasing the recognition accuracy up to 87%.