Multimodal Learning-Driven Software Analysis

随着软件系统在真实的世界生产环境中继续扮演越来越重要的角色，它们的可靠性变得越来越重要。软件错误每年使全球经济损失超过1万亿美元，并可能造成毁灭性的破坏和人类生命。软件开发人员花费大约50%的工作时间来查找和修复与软件相关的错误。该研究计划的目标是自动减轻，检测和修复错误的新软件分析技术。我们正处于一个关键点，我们可以使用深度学习从大量数字数据中自动学习模式。这使我们能够解决一些经常出现的软件工程挑战，例如错误检测和修复，通过对从软件开发人员那里收集的数据进行深度学习。传统的自动错误检测和程序修复技术依赖于一组预定义的模板和规则，这些模板和规则仅限于特定的软件错误类型;添加对新类型错误的支持是手动的、临时的和昂贵的。与硬编码错误检测和修复规则不同，我们可以自动地从开发人员犯错误的方式中学习并修复这些错误。这里的挑战是如何最好地表示适合机器学习的矢量化软件系统。 软件分析传统上围绕着软件的源代码。我们建议采取多模态学习方法来进行错误检测和修复，其中除了软件的语法和语义之外，还考虑到其运行时行为，通过软件分析，计算机视觉和机器学习算法的组合。我们将设计新的技术，用于对程序执行痕迹和运行时视觉人工的文本和视觉分析。这项在软件分析、深度学习和计算机视觉交叉领域的拟议计划将支持我们在自动错误检测和修复方面的世界领先举措，并有助于进一步确立加拿大在这一重要而及时的研究领域的领导地位。这项工作将为采用和应用支持机器学习的软件的企业带来直接的下游利益，在部署到现实环境之前检测和修复错误。我们相信，我们的多模态学习方法将对设计更准确的软件分析技术产生重大影响，不仅适用于错误检测和修复，而且适用于软件工程的许多相关领域，其中任务可以通过从示例中学习来自动化。

As software systems continue to take on ever more central roles in real--world production settings, their dependability has become increasingly critical. Software errors cost the global economy over $1 trillion per year and can cause devastating disruptions and human lives. Software developers spend around 50% of their work time finding and fixing software- related errors. This research program targets novel software analysis techniques for mitigating, detecting, and repairing errors automatically. We are at a pivotal point in which we can automatically learn patterns from massive amounts of digital data using deep learning. This enables us to tackle some of the recurring software engineering challenges, such as error detection and repair, by applying deep learning on data gathered from software developers in practice. Traditional automated error detection and program repair techniques rely on a set of predefined templates and rules that are limited to specific software error--types; adding support for a new type of error is manual, ad--hoc, and costly. Instead of hard--coding error detection and repair rules, we can automatically learn from the way developers make mistakes and repair those mistakes. The challenge here is how to best represent software systems for vectorization amenable to machine learning. Software analysis has traditionally revolved around the source code of the software. We propose to take a multimodal learning approach to error detection and repair in which in addition to the syntax and semantics of the software, its runtime behaviour is taken into account, through a combination of software analysis, computer vision, and machine learning algorithms. We will devise new techniques for textual and visual analysis of program execution traces and runtime visual artificants. This proposed program at the intersection of software analysis, deep learning, and computer vision will bolster our world- leading initiatives in automated error detection and repair and help further establish Canada as a leader in this important and timely research area. This line of work will have direct downstream benefits for businesses that adopt and apply machine learning-enabled software by enabling the detection and fixing of errors before deployment to real-world settings. We believe our multimodal learning approach will have a significant impact in devising more accurate software analysis techniques, not only for bug detection and repair, but for many related areas of software engineering where tasks can be automated by learning from examples.