I am a research intern in the BrAIN group at the Max Planck Institute for Software Systems (MPI-SWS) advised by Prof. Mariya Toneva. My primary research interests are centered around the intersection of machine learning and computational neuroscience.
Previously I was a (research) project assistant at the Indian Institute of Science under the guidance of Prof. Anirban Chakraborty, where I developed methods for training adversarially robust neural networks under low-resource constraints.
Feel free to drop me an email, if you would like to discuss possible collaborations!
Performance of existing certified defenses significantly deteriorates as training data size is reduced. We introduce a novel data-augmentation technique that
improves their data-efficiency.
Gaurav Kumar Nayak*, Ruchit Rawal*, Anirban Chakraborty.
|
|
Dynamics of Dataset Bias and Robustness
Principles of Distribution Shift (PODS), ICML-Workshop 2022
Exploring the effects of various techniques for improving robustness under distribution-shift on the dataset-bias.
Prabhu Pradhan*, Ruchit Rawal*
|
|
Holistic Approach To Measure Sample-Level Adversarial Vulnerability and Its
Utility in Building Trustworthy Systems
CVPR-2022 Wokrshop on Human-centered Intelligent Services: Safety and Trustworthy
Propose a holistic approach for quantifying adversarial vulnerability of a sample by combining multiple perspectives.
Gaurav Kumar Nayak*, Ruchit Rawal*, Rohit Lal*, Himanshu Patil, Anirban Chakraborty.
|
|
We propose a novel problem of "test-time adversarial defense in absence of training data and even their statistics" and provide first steps to solve it.
Gaurav Kumar Nayak*, Ruchit Rawal*, Anirban Chakraborty.
|
|
Learning modality invariant features is central to the problem of Visible-Thermal cross-modal Person Reidentification (VT-ReID). We propose a simple but effective
framework, MMD-ReID, that reduces the modality gap by an explicit discrepancy reduction constraint.
Chaitra Jambigi*, Ruchit Rawal*, Anirban Chakraborty
|
|
Rendezvous between Robustness and Dataset Bias: An empirical study
NeurIPS-2020 pre-registration Workshop
We shine a light on the effects of various techniques for perturbarial robustness on dataset-distribution bias (i.e. class imbalance).
Prabhu Pradhan*, Ruchit Rawal*, Gopi Kishan
|
|
Generalizing across the (in)visible spectrum
ICML-2020 Workshop on Extreme Classification, ORAL
Improvements in robustness do not translate to better intrinsic generalization (esp. in case of high imbalance or long-tail scenarios).
Ruchit Rawal*, Prabhu Pradhan*
|
|
Climate Adaptation: Reliably Predicting from Imbalanced Satellite Data
CVPR-2020 Workshop on AGRICULTURE-VISION, ORAL
Real-World datasets usually suffer from high class imbalance, impeding a model's ability to learn effectively. We provide a comprehensive guide on the effectiveness of various state-of-the-art methodologies in a combination setting.
Ruchit Rawal*, Prabhu Pradhan*
|