Principal Computer Scientist,
Computer Science Laboratory,
Email: susmit.jha at sri.com
, jha at eecs.berkeley.edu
Staff Research Scientist, UTRC, Berkeley (Dec, 2014 - Oct, 2016)
Research Scientist, Intel, Hillsboro (Jan, 2012 - Dec, 2014)
M.S. and Ph.D., EECS, UC Berkeley (Aug, 2006 - Dec, 2011)
B.Tech., Dept. of Computer Science, IIT Kharagpur (2002-2006)
Susmit Jha is a Principal Scientist at SRI International
, leading the research thrust on trust, resilience, and interpretability of AI.
His research focuses on
combining formal methods and machine learning to build
trusted artificial intelligence and correct-by-construction autonomous
systems. His research background is in Formal Methods, Artificial Intelligence and Control Theory. His current projects are:
: We have been selected to perform on DARPA IDAS
on the use of symmetry for automated synthesis, and IARPA TrojAI
for detecting Trojans/backdoors in deep learning models.
Previous Projects: DARPA SW Symmetry, DARPA RFMLS
, DARPA BRASS
Towards Automated System Synthesis Using Sciduction
, UC Berkeley, November 2011
: Sanjit A. Seshia
, Rest of Thesis Committee
: Claire Tomlin
, Dorit Hochbaum
: This thesis combined learning-based induction and logical deduction to automate synthesis of programs and controllers from examples and demonstrations.
: Automated synthesis of systems that are correct by construction has been a long-standing goal of computer science. Synthesis is a creative task and requires human intuition and skill. Its complete automation is currently beyond the capacity of programs that do automated reasoning. However, there is a pressing need for tools and techniques that can automate non-intuitive and error-prone synthesis tasks. This thesis proposes a novel synthesis approach to solve such tasks in the synthesis of programs as well as the synthesis of switching logic for cyberphysical systems. The common underlying theme of the proposed synthesis techniques is a novel combination of deductive reasoning, inductive reasoning and structure hypotheses on the system under synthesis. We call this combined reasoning technique SCIDUCTION that stands for ‘Structurally Constrained Induction and Deduction’. SCIDUCTION constrains inductive and deductive reasoning using structure hypotheses, and actively combines inductive and deductive reasoning: for instance, deductive techniques generate examples for learning, and inductive techniques generate generalizations as candidate designs to be proved or disproved by deduction. We use the proposed synthesis approach for automated synthesis of loop-free programs from black-box oracle specifications using functions from a library of component functions, synthesizing optimal cost fixed-point code with specified accuracy from floating-point code, and synthesizing switching logic of hybrid systems for safety and performance properties. We illustrate that our approach can be used to automate system synthesis, and thus, can prove to be an effective aid to designers and developers.
Primary thesis papers
: Leon Chua Award, UC Berkeley, 2011 for contributions to nonlinear sciences.