About Me
I am a final-year PhD student in Computing and Information Sciences at Rochester Institute of Technology, advised by Prof. Linwei Wang in the Computational Biomedicine Lab. My research focuses on making deep generative models more reliable: improving their robustness and controllability for cases underrepresented in the training data.
My current work explores these themes in computer vision:
- Robustness to attribute correlations. I study how attribute correlations in training data make it difficult for generative models to produce images with underrepresented attribute combinations, and explore approaches to addressing this challenge.
- Controllable and counterfactual generation. Going a step further, I develop causal generative models that account for relationships between attributes to improve controllable and counterfactual generation accuracy under varying strengths of these relationships.
Selected Recent Research
View all publications →ProReGen: Progressive Residual Generation under Attribute Correlations
Correlations among attributes in training data can cause generative models to struggle with requested combinations that conflict with those correlations. ProReGen addresses this through attribute residualization and progressive learning: it first learns from abundant, correlation-aligned samples, then uses limited correlation-conflicting samples to learn residual information. The framework improves attribute correctness for underrepresented combinations across VAEs, GANs, and diffusion models.
Education
Rochester Institute of Technology
Tribhuvan University, Nepal