I lead the Davis Automated Reasoning Group (DARG), a team of researchers developing tools and techniques for ensuring correctness of software. The use of software is prevalent in our daily lives, and is increasingly used in safety and security-critical systems. Software systems come in a variety of shapes and sizes, from firmware for IoT devices consisting of thousands of lines of code to operating systems consisting of millions of lines of code to deep neural networks with billions of parameters. Each such system has different performance, security, safety, and functional requirements. Our research develops theoretical formalisms as well as practical tools to tackle these challenges, with the goal of developing foundational research and transitioning it to industry to have real-world impact.
A short description of some of our research can be found below; more details can be found in the publications. Our software tools are available on GitHub. I discussed some of our projects as well as my research philosophy in Episode 7 of the Build Better Systems Podcast; interview was in Aug. 2020.
Analysis of Deep Neural Networks
Deep neural networks (DNNs) have become the state-of-the-art in a variety of applications including image recognition and natural language processing. Moreover, they are increasingly used in safety and security-critical applications such as autonomous vehicles. This project aims to improve our understanding of DNNs by developing techniques to analyze, verify, and repair them.
Efficient Abstract Interpretation
Abstract interpretation is a general framework for expressing static program analyses. This project aims to improve the time and memory performance of abstract interpretation, specifically the fixpoint computation central to all abstract interpreters. Our research is used in two open-source abstract interpreters: NASA IKOS and Facebook SPARTA.
This video from the Workshop on Research Highlights in Programming Languages at FSTTCS 2020 describes some our results.
Neuro-Symbolic Program Analysis
Neuro-symbolic program analysis augments static program analysis with machine learning to find bugs in large systems software, such as the Linux kernel.