CAREER: Enhancing Deep-Learning-based Code Analyses via Human Intelligence

该奖项全部或部分由《2021年美国救援计划法案》（公法117-2）资助。随着开源存储库中数以百万计的程序的可用性不断增加，人们提出了许多技术来利用深度学习模型从大型代码库中自动学习模式，以协助各种软件工程任务（例如，安全分析，错误检测）。然而，由于许多原因（例如，程序代码的演变），所提出的深度学习模型仍然面临许多超出模型处理能力的输入程序。由于缺乏对这些输入的理解，工业实践中的许多软件工程应用仍然基于符号推理系统做出决策，其中决策逻辑和规则是由人类硬编码的。人类智能（例如，人类总结的规则）往往是简单和简化的，而深度学习模型可能是不透明和过拟合的。如果一个人能以某种方式将这两个世界的优点结合起来，那么许多现有的挑战就会消失。因此，本建议力求在这种结合方面取得进展。该提案的总体目标是设计一个通用框架，通过结合人类智能来改进深度学习模型对输入程序的处理。具体而言，现有深度学习模型在处理代码数据时面临两个主要问题：(1)缺乏对代码数据固有性质的理解；(2)缺乏对软件工程任务的特定领域知识。为了解决这些基本限制，本项目建议设计一个通用的、用户驱动的、基于学习的框架。短期内，该项目旨在提高智能代码分析技术的实用性，并促进深度学习技术在代码分析中的应用。从长远来看，这个项目有潜力从根本上改变软件工程应用程序中基于学习的代码分析技术。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).With the increasing availability of the millions of programs in open-source repositories, many techniques have been proposed to leverage deep-learning models to automatically learn patterns from large code bases to assist various software engineering tasks (e.g., security analysis, bug detection). However the proposed deep-learning models still face many input programs that are beyond a model’s handling capability due to many reasons (e.g., evolution of the program code). Because of the lack of understanding about these inputs, many software-engineering applications in industrial practice still make decisions based on symbolic-reasoning systems where decision logic and rules are hard-coded by a human. Human intelligence (e.g., rules summarized by humans) tends to be simplistic and reductionistic, while deep-learning models can be opaque and overfitted. If one can somehow combine the best of the two worlds, many existing challenges will disappear. Therefore, this proposal seeks to make progress on such a combination. The broad goal of this proposal is to design a general framework that improves deep-learning models’ handling of input programs by incorporating human intelligence. Specifically, two main issues are faced by existing deep-learning models in handling code data: (1) lack of understanding about inherent nature of code data, and (2) lack of domain-specific knowledge of software-engineering tasks. To address these fundamental limitations, this project proposes to design a general, user-driven learning-based framework. In the short term, this project aims to improve the practicality of intelligent code-analysis techniques and facilitate the adoption of deep learning techniques in code analysis. In the long run, this project has the potential to fundamentally transform the learning-based techniques for code analysis in software-engineering applications.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.