APPCAIR-I / Ataxia-Sync
Developed a complete pipeline around Spatio-temporal GNNs capable of
diagnosing Ataxia in patients and it's severity,
runs on low compute devices.
Out on arXiv.
More details after the review process.
APPCAIR-II / Radio-LM
The project revolves around finding harmful / beneficial aspects of LLMs and their explanations.
We are testing this using Radiology/Med students.
This has been studied in the past by Donald Michie who proposed a "human-window" of task-doability.
This work's presentation led to the first prize in ACM event, slides.
This work also takes inspiration from the PXP protocol (arXiv),
for which, I have co-developed an implementation GitHub.
DeepForestSciences
Contributing to the DeepChem library.
I am actively developing the Materials Science section; as a part of which I have edited to the
Introduction to Materials Science Tutorial.
I am adding several more basic-equivariance tutorials, as well as developing a pipeline for arbitrary property optimization in Materials Science
(along with Applied Materials!).
Previously, I co-developed MolGAN and Normalizing Flow implementations in PyTorch.
This work (arXiv) has been accepted at
MoML'24,
BayLearn'24,
and also ML4PS @ NeurIPS'24.
I also developed the tutorials for them,
MolGAN,
NormFlows.
[RE] Teaching CLIP to Count to Ten
This is a student-led reproduction of the Google Brain paper Teaching CLIP to Count to Ten. We developed novel (incremental) augmentations to the loss function, which showed improvement in our small scale experiments. Apart from this, we created data processing pipelines and also made their evaluation dataset completely public. This work has been submitted to ReScience C, arXiv.
Event Stream analysis using Deep SSMs
This project's report is under review at a double-blind venue. Details soon.