Mukilan Karuppasamy

I'm a predoctoral researcher at IISc, Bangalore, working with Anirban Chakraborty on visual instruction tuning for multimodal large language models. Previously, I had the opportunity to work with CV Jawahar at IIIT Hyderabad on concept-based interpretability in video action recognition, and with Yogesh Singh Rawat at CRCV, UCF, on robustness analysis of vision-language models. I was also a MITACS Globalink Fellow in 2023, working on distributed message passing with Amine Mezhghani and Faouzi Bellili at the University of Manitoba. Earlier, I worked on virtual try-on systems using GANs with Brejesh Lall at IIT Delhi.

Email  /  CV  /  Scholar  /  Github  /  Linkedin

Research

I'm interested in computer vision, interpretability, and robustness analysis. My current interest lies in understanding the theoretical basis of interpretability in overparameterized deep neural networks. Some papers are highlighted.

Proposed a concept based inherently intrepretable model for Video models using concept bottleneck model and token merging. This was applied in Driver Intention Prediction task.

Derived a collaborative signal estimation from multiple agents with different measurement channels through distributed message passing algorithm.

We critically examine the vulnerability of SOTA Zero-Shot Object Detectors with 50 different backbones to resolution degradation (`pixelation'). Two key takeaways are 1) The choice of backbone alone can substantially enhance robustness, while prompting (language) plays a minimal role. 2) Robustness is also a function of the dataset/domain of images at inference. A certain type of image are not affected by pixelation.

Proposed a large scale sketch based pretraining and visual instruction tuning dataset with 30M instances. Developed a state of the art Sketch based Multimodal model which can perform detection, Counting, VQA, SBIR with any crude, hand drawn sketches as the query instead of natural language.

Services

Reviewer - WACV 2026, AAAI 2026.