Research Grants

Novel AI methods for Software Debugging

Aditya Kanade and Shirish Shevade

Software Debugging is one of the most time consuming activities forprogrammers. Automatic fixing of programming errors (program repair) has therefore become an important research topic in Software Engineering in recent years. Traditionally, program analysis techniques based on programming language semantics have been designed to aid in software debugging. However, scaling them to large codebases and maintaining low rate of false positives requires significant and continuous manual efforts. This has limited their applicability in practice. We propose to take advantage of availability of large codebases (such as student programs in MOOCs and open-source git repositories) and associated bug reports/fixes to develop novel artificial intelligence (AI) approaches to aid developers in software debugging.

Causal Inference for Human Brain Networks

Arnab Bhattacharyya and Sridharan Devarajan

Traditional statistical inference treats all observables on an equal footing, ignoring the fact that reality inevitably imposes directed relationships. In fact, causal relations can be impossible to deduce from purely statistical considerations alone: correlation does not imply causation. Causal mechanisms can be uncovered by interventions, where an intervention consists of setting a subset of variables to fixed values and observing the effect on the rest. The overall goal of our work is to recover from interventional data a causal network which displays how the variables are causally dependent on each other, while minimizing the number of interventions and the number of test samples for each. We address some basic quantitative questions in this direction.

We will then apply this theoretical approach for the estimation of causal networks from brain imaging data. Specifically, we will record brain activity along with applying appropriate interventions, by focally manipulating the activity of candidate brain nodes. These interventions will be numerically tested with computational simulations of fMRI brain activity, and experimentally validated with non-invasive brain stimulation techniques (like transcranial magnetic stimulation). The results may provide key insights into identifying malfunctioning brain networks for early diagnosis and treatment of neurological disorders.