Light-Weight SA-BNN: High-Precision Olfactory Recognition of Binary Neural Networks

Yijun Qin, Huajun Ma, Shukai Duan, Lidan Wang

Abstract: With the development of artificial intelligence, electronic nose technology, which simulates human olfaction using sensors and pattern recognition techniques, is becoming increasingly mature and widely applied in various fields. Currently, the application of lightweight techniques to electronic noses makes it challenging to achieve high precision. This paper proposes a Binary Neural Networks (BNNs) based on knowledge distillation for electronic nose recognition. We present our own BNNs model, and due to the performance gap between BNNs and Full Precision Neural Networks (FNNs), we introduce the Self-Attention (SA) mechanism to enhance accuracy in this model. Knowledge distillation, a training method based on a student-teacher network, is also introduced to achieve model lightweight. By comparing network models with different weights, we demonstrate that binary network models are more suitable for odor recognition. Experimental results show that SA-BNN achieves superior performance with a 99.64% accuracy on the UCI dataset compared to other networks. Binary computation reduces computation and storage requirements, and through knowledge distillation, the model parameters are reduced by 2/3, achieving an accuracy of 97.76%, making it more suitable for embedded devices. This study lays the foundation for the lightweight application of biological neural networks in odor recognition and has the potential for edge computing.