Lightweight Binary Neural Networks for Edge Devices

The necessity to operate artificial neural networks on edge devices and mobile devices efficiently with limited resources is becoming increasingly prominent, and Binary Neural Networks (BNNs) are emerging as an alternative to this. In this paper, we propose an improved BNN model using Vitis HLS that can significantly reduce resource usage and power consumption at the expense of a slight reduction in accuracy. By applying three types of optimization methods proposed in this paper, resource usage and power consumption are significantly reduced by up to $93 \%$ and $30 \%$, respectively, compared to the conventional model. This effectively demonstrates the possibility of implementing a high-performance deep learning model even in a resource-limited environment.