XNOROP: a New Metric for Binary Neural Networks Performance Measurement

Abstract

Abstract—The quantity of multiply–accumulate operations in a model (MACs) is a widely used metric in the field of neural networks and deep learning. This metric expresses the computational cost of a considered model. However, in Binary Neural Networks BNNs, we merely have floating point operations at inference time, we have mostly XNOR binary operations, hence we can’t use the metric MACs to describe the computation cost of BNNs. In this paper, we propose XNOROPs; a new metric that expresses the computation cost of BNNs on Central Processing Units CPUs and Microcontrollers unit MCUs. A compression technique for BNNs is introduced to maximize the usage of CPU and MCU resources, and so reduce the inference time. The new metric is well explained and built-up step by step taking into consideration the inner operations cost in terms of CPU and MCU cycles. XNOROP is related to the well-known MAC metric by a mathematical equation, enabling us to measure the number of operations in a binarized model when number of operations in its float counterpart is provided. Finally, a recipe that helps is choosing the appropriate hardware for BNNs deployment using the new XNOROP metric is provided.

Keywords

• binary neural networks
• flops
• xnorops
• bnn compressing
• bnn metric