Testing, Debugging and Repairing Machine Learning Software at the System Level

点击翻译按钮获取中文摘要

Machine learning (ML) has become the driving force of innovation in many application domains. However, current state-of-the-art ML software still suffers from quality issues. Different from traditional software, ML software adopts the data-driven programming paradigm. Instead of manually programming the decision logic (e.g., in form of source code), the decision logic of the ML software is automatically learned via training and encodes in a model (e.g., neural network). An ML model is often difficult to interpret and understand, calling for novel quality assurance (QA) methods. In practice, an ML model is often not used standalone but integrated as a component into a larger system, including both traditional software and ML models. While some recent progress is made on testing and analysis ML models, systematic research of QA for ML software at the system level is still largely untouched so far. Quality assurance at the ML system level is challenging and requires considering the components of ML models, traditional software, and their interactions. In particular, 1) The ML model behaviors are often difficult to understand. What is the role of an ML model and how its incorrect behaviors impact the whole system? 2) How to effectively detect the defects in the huge testing space of an ML software system? 3) With the error triggering tests, how to debug and identify the defect modules? 4) For system incorrect behaviors caused by ML models, how to repair them to improve system quality? This research program aims to address these challenges and propose novel methods of testing, debugging, and repairing for ML-driven software at the system level, providing key quality assurance supports to establish trustworthy intelligent software. 1) First, I plan to perform a large-scale empirical study to systematically investigate roles and defect impacts of ML models in state-of-the-art ML systems. 2) Then, I will propose an effective testing framework to detect the potential defects of ML at the system level. 3) With the found defects, I plan to design automated debugging techniques to localize the potentially incorrect modules. 4) Regarding the system defects introduced by ML models, I will further propose automated repairing methods to enhance the ML system quality. Large-scale experiments on open source and industrial ML software systems will be conducted to evaluate the advantage, practical value and limitation of proposed techniques. This outcome of this research will originally provide an initial set of key methods to detect, debug, and repair ML software at the system level, which can greatly accelerate the ML system development process with better quality assurance support, potentially impacting many industrial domains. This program will train nine highly qualified personnel (HQP) and provide them with the equity, diversity and inclusivity (EDI) platform to participate and contribute to the state-of-the-art intelligent software engineering research.