Characterization of binarized neural networks for efficient deployment on resource-limited edge devices

Abstract

This paper delves into binarized neural networks (BNNs) tailored for resource-constrained edge devices. BNNs harness binary weights and activations to amplify efficiency while upholding accuracy. Across diverse network configurations, BNNs consistently outshine traditional neural networks (NNs). A pioneering BNN architecture is developed in LARQ, achieving an impressive. 61% accuracy on the MNIST dataset through binary quantization, weight clipping, and pointwise convolutions. Implementation on the Xilinx PYNQZ2 FPGA board shows far quicker classification rates, with a maximum inference time of 0.00841 milliseconds per image, approximately 10,000 images being classified in this length of time. The time taken per image represents approximately 0.01% of the total inference time. This underscores BNNs' potential to redefine real-time edge computing applications. The paper makes significant strides by elucidating BNNs' performance superiority, proposing an innovative architecture, and validating its prowess through real-world deployment. These findings underscore BNNs as agile, high-performance models primed for edge computing, fostering a new era of real-time processing innovations.