Research
I am interested in understanding the failure modes of the Machine Learning model and using these insights to build a more robust and generalizable model. So far, I have looked at the problem from a causal perspective using tools from Causal Inference and Causal Discovery. I am also interested in looking into this problem from an optimization and robust learning perspective.
|
|
Probing Classifiers are Unreliable for Concept Removal and Detection
Abhinav Kumar,
Chenhao Tan,
Amit Sharma
NeurIPS, 2022
arXiv
/
NeurIPS Reviews
/
ICML SCIS Workshop
/
Poster
/
Slides
We show that latent space based concept detection and removal methods like Null-Space Removal (INLP) and adversarial removal, which internally uses probing classifiers, are unreliable. They fail to remove the desired concept and, in the worst case, remove or corrupt other features or concepts from the latent space of the classifier.
|
|
Disentangling mixtures of unknown causal interventions
Abhinav Kumar,
Gaurav Sinha
UAI, 2021
Conference Paper
/
arXiv
/
Poster
/
Slides
/
Video
We study the problem of disentangling a mixture distribution, where components of the mixture came from interventional distributions on a base causal graph.
|
|