Deep Neural Networks (DNN) represent a performance-hungry application. Floating-Point (FP) and custom floating-point-like arithmetic satisfies this hunger. While there is need for speed, inference in DNNs does not seem to have any need for precision. Many papers experimentally observe that DNNs can successfully run at almost ridiculously low precision.
The aim of this paper is two-fold: first, to shed some theoretical light upon why a DNN's FP accuracy stays high for low FP precision. We observe that the loss of relative accuracy in the convolutional steps is recovered by the activation layers, which are extremely well-conditioned. We give an interpretation for the link between precision and accuracy in DNNs.
Second, the paper presents a software framework for semi-automatic FP error analysis for the inference phase of deep-learning. Compatible with common Tensorflow/Keras models, it leverages the frugally-deep Python/C++ library to transform a neural network into C++ code in order to analyze the network's need for precision. This rigorous analysis is based on Interval and Affine arithmetics to compute absolute and relative error bounds for a DNN. We demonstrate our tool with several examples.
Semi-Automatic Data Annotation guided by Feature Space Projection
Institute of Computing, University of Campinas（坎皮纳斯大学计算研究所）；Department of Information and Computing Sciences, Faculty of Science, Utrecht University（乌得勒支大学理学院信息与计算科学系）
ReconVAT: A Semi-Supervised Automatic Music Transcription Framework for Low-Resource Real-World Data
Institute of High Performance Computing (IHPC), Agency for Science, Technology and Research (A*STAR)（科学、技术与研究署（A*STAR）高性能计算研究所（IHPC））；Singapore University of Technology and Design（新加坡科技设计大学）
