I am an Assistant Professor at the Yardi School of Artificial Intelligence, IIT Delhi. 

Prior to this, I was a postdoctoral associate at Machine Learning Department, MBZUAI, where I worked with Prof. Bin Gu and Prof. Huan Xiong. I obtained a Ph.D. in Machine Learning from the Department of CSE, IIT Kanpur, under the supervision of Prof. Purushottam Kar and Prof. Sandeep K. Shukla.

My work can be broadly categorized into the following  areas:
Training-time Robustness: As ML models are trained on increasingly larger and untrusted data,  the chances that part of the training data is corrupted become real. Designing training algorithms that perform reasonably well (if possible with provable guarantees) despite training data corruption is a form of training-time robustness, addressed in some of our works [AAAI 2023, MachLearn 2021, AISTATS 2019].

Test-time Robustness: Further, as ML models are deployed in real environments (e.g., self-driving cars, face recognition software, Language Model APIs), they can be exposed to inputs that are deliberately modified to alter the model's required behavior. Some of our work shows that a model is provably robust to adversarial input perturbations, with minimal compromise in the model's standard behavior [ICLR 2025, ICLR 2024].

Energy-Efficient Computing: The human brain consumes far less energy than a modern computer simulating far fewer neurons. A key reason for this is the spiking behavior of biological neurons, in which very few neurons in the network produce an output when presented with a particular input. In contrast, all neurons in a standard artificial neural network must process a particular input signal. Spiking neural networks are an energy-efficient computing model that emulates this biological behavior. Some of our work has examined efficient training and inference for such models [AAAI 2024, NeurIPS 2023], as well as their robustness properties mentioned earlier.