Deep learning (DL) has been applied widely, and the quality of DL system becomes crucial, especially for safety-critical applications. Existing work mainly focuses on the quality analysis of DL models, but lacks attention to the underlying libraries and frameworks on which all DL models depend. In this work, we propose Audee, anovel approach for testing DL libraries and localizing bugs. Audee adopts a search-based approach and implements three different mutation strategies to generate diverse tests cases by exploring combinations of model structures, parameters, weights and inputs. Audee is able to detect three types of bugs: logic bugs, crashes and Not-a-Number (NaN) bugs. In particular, for logic bugs, Audee adopts a cross-reference check to detect behavioral inconsistencies across multiple frameworks (e.g., TensorFlow and PyTorch), which indicates potential bugs in their implementations. For NaN bugs, Audee adopts a heuristic-based approach to generate DNNs that tend to output outliers (i.e., too large or small values), and these values are likely to cause NaN value. Furthermore, Audee leverages causal testing based technique to localize layers as well as parameters that cause inconsistencies or bugs. To evaluate the effectiveness of our approach, we applied Audeeon evaluating four DL frameworks, i.e., TensorFlow, CNTK, Theano, and PyTorch. We totally generate 260 models which cover 25 widely-used APIs in the four frameworks. The results demonstrate Audee are effective indetecting inconsistencies, crashes and NaN bugs. In total, 26 unique unknown bugs were discovered, and seven of them have already been confirmed by the developers.

Qianyu Guo

College of Intelligence and Computing, Tianjin University

China

Xiaofei Xie

Nanyang Technological University

Yi Li

Nanyang Technological University, Singapore

Singapore

Xiaoyu Zhang

Xi'an Jiaotong University

Yang Liu

Nanyang Technological University, Singapore

Singapore

Li Xiaohong

TianJin University

China

Chao Shen

Xi'an Jiaotong University

Neural networks are becoming a popular tool for solving many real-world problems such as object recognition and machine translation, thanks to its exceptional performance as an end-to-end solution. However, neural networks are complex black-box models, which hinders humans from interpreting and consequently trusting them in making critical decisions. Towards interpreting neural networks, several approaches have been proposed to extract simple deterministic models from neural networks. The results are not encouraging (e.g., low accuracy and limited scalability), fundamentally due to the limited expressiveness of such simple models.

In this work, we propose an approach to extract probabilistic automata for interpreting an important class of neural networks, i.e., recurrent neural networks. Our work distinguishes itself from existing approaches in two important ways. One is that probability is used to compensate for the loss of expressiveness. This is inspired by the observation that human reasoning is often `probabilistic’. The other is that we adaptively identify the right level of abstraction so that a simple model is extracted in a request-specific way. We conduct experiments on several real-world datasets using state-of-the-art architectures including GRU and LSTM. The result shows that our approach significantly improves existing approaches in terms of accuracy or scalability. Lastly, we demonstrate the usefulness of the extracted models through detecting adversarial texts.

Guoliang Dong

Computer College of Zhejiang University

Jingyi Wang

Zhejiang University

Singapore

Jun Sun

Singapore Management University

Singapore

Yang Zhang

Zhejiang University

Xinyu Wang

Zhejiang University

China

Dai Ting

Huawei International Pte Ltd

Jin Song Dong

National University of Singapore

Xingen Wang

Zhejiang University

Data augmentation techniques that increase the amount of training data by adding realistic transformations are used in machine learning to improve the level of accuracy. Recent studies have demonstrated that data augmentation techniques improve the robustness of image classification models with open datasets; however, it has yet to be investigated whether these techniques are effective for industrial datasets. In this study, we investigate the feasibility of data augmentation techniques for industrial use. We evaluate data augmentation techniques in image classification and object detection tasks using an industrial in-house graphical user interface dataset. As the results indicate, the genetic algorithm-based data augmentation technique outperforms two random-based methods in terms of the robustness of the image classification model. In addition, through this evaluation and interviews with the developers, we learned following two lessons: data augmentation techniques should (1) maintain the training speed to avoid slowing the development and (2) include extensibility for a variety of tasks.

Haruki Yokoyama

Fujitsu Laboratories Ltd.

Satoshi Onoue

Fujitsu Ltd.

Shinji Kikuchi

Fujitsu Laboratories Ltd.