I am a research scientist at Bosch Research (Sunnyvale, USA), where I work on computer vision problems for autonomous driving. My research interests are model interpretability and the “science” of deep learning, i.e., systematic investigations of deep learning phenomena.

I started my research career with a masters at the Indian Institute of Science, Bangalore with Prof. Venkatesh Babu. I completed my PhD with François Fleuret, at Idiap Research Institute & EPFL, Switzerland. Enamoured by model interpretability, I pursued a postdoctoral research fellowship with Hima Lakkaraju at Harvard University.

Some representative papers include:

• splice: shows that CLIP representations are approximate sparse linear combinations of concept vectors, using this to develop an interp tool that extracts concepts encoded by representations
• explaining perceptually aligned gradients: an experimental theory explaining why noise-robust models tend to produce improved (perceptually aligned) gradient saliency maps
• fullgrad saliency: shows that ReLU neural net outputs can be exactly decomposed into layer-wise gradient terms, using this to develop a saliency tool that aggregates layer-wise gradients
• forgetting data contamination in LLMs: an experimental study showing that data contamination in LLMs may be irrelevant in certain practical training scenarios

For more information, please see my research themes and publications.

Note: If you’d like to chat about the science or interpretability of neural nets, feel free to reach out!