I am a postdoctoral research scientist at Columbia University, where I am advised by Carl Vondrick. I did my PhD from Cornell University co-advised by Kavita Bala and Bharath Hariharan. Prior to that, I obtained my bachelor’s degree in Computer Science and Engineering from Indian Institute of Technology Bombay.
My research interest lies in computer vision and its application. More specifcally, I aim to build recognition models that can learn with little to no supervision. I also use these models to make discoveries and provide scientific insights from visual data in various scientific domains. I have applied my work to a range of application domains from fashion to satellite (remote sensing) images.
Note: If you are an undergrad or masters student at Columbia or Cornell and are interested in Vision for Science Research, reach out to me for potential project opportunities.
Here is a list of my publications:
TL;DR: A generative domain translation method that identifies and reports spatio-temporal features distinguishing facial expressions in different communication contexts, enhancing understanding of behavioral differences.
TL;DR: An evolutionary search algorithm using LLMs to iteratively discover interpretable and discriminative classifiers for visual recognition.
TL;DR: A vision-language model for satellite images, trained by using geo-located internet images as intermediary between text and satellite images.
TL;DR: A self-supervised representation learning approach for satellite images that uses temporal and change information to learn better representation.
TL;DR: A method to create benchmarks for discovering meaningful multi-step change events from satellite images with no labels.
TL;DR: A practical improvement on zero-shot learning, allowing annotators to provide multiple descriptors for a concept with multiple modes of appearance.
TL;DR: A method to discover neighborhood similarity in a city using the fashion characteristics withing a city.
TL;DR: A practical active-learning interface to efficiently specify attributes in zero-shot learning.
TL;DR: An unsupervised semantic segmentation model by clustering and encouraging equivariance to geometric transforms and invariance to photometric ones.
TL;DR: An automated framework analyzing fashion from street photos for accurate forecasting of fashion trends/style and discovering social/cultural and sporting events.
TL;DR: A method offering a potentially tighter bound on iterations compared to previous variants of Policy Iteration (PI) algorithms.
TL;DR: Quantitative comparison of how mainstream clothing retail brands represent model skin tones across still and video media modes.
TL;DR: A tool to analyze large datasets of fashion imagery, revealing trends of fine-grained concepts such as baseball caps.
Advisor: Carl Vondrick
Advisor: Kavita Bala and Bharath Hariharan
Minor in Cognitive Science
Advisor: Siddhartha Chaudhuri
Minor in Bio-sciences and Bio-engineering