Floating Point Accuracy Testing in Deep Neural Network Computations via Hypothesis Testing
Chaojin Wang, Jian Shen · 2020
Deep Learning (DL) has shown its success and convenience in our daily life, especially for the application in mobile devices. Because of the high quality demand for mobile applications from the consumers, the floating point accuracy of computations especially in deep neural networks (DNNs) has become particularly important. In this paper, we focus on testing the accuracy of floating point in DNN computations in the mobile devices from a statistical point of view. Specifically, two common hypothesis testing methods Z- Test and Student's T- Test are utilized, which cope with different situations. Z- Test is suitable for samples with known population variance and T-Test is suitable for samples without known population variance. Compared to existing scheme that pays more attention on single value comparison in industry, our proposed scheme in field of probability theory is more representative and reaches a more credible results with quantitative measurement, which can be seen from our experiment that based on well-known accelerator HiAI. The most important contribution of this paper is a novel perspective to test floating point accuracy. Future work is to study how to improve the accuracy based on the results.