Karan Bania

(કરન બનિયા)
Karan Bania
I'm a final year Computer Science undergraduate at BITS Goa, India, currently doing my bachelor's thesis at Gan.AI. Previously, I have interned at APPCAIR and DeepForestSciences. I was also a DAAD WISE fellow with ESML for Summer'24.

I really like rap music [rap&blues] and am an avid Pokémon catcher ;). Sometimes, I also (try to) talk about research, here's the better ones, Graphormer [Slides / BlogPost], Radio-LM, Deep SSMs [Slides / Video] and Wav2Lip [Slides / Video].

Quite amusingly, I have an Erdős number of 4. Paul Erdős -> Endre Szemerédi -> Wolfgang Maass -> Anand Subramoney -> Karan Bania

[Email] [CV] [GitHub] [Medium] [Scholar]



News



Experience

Research Engineer Intern
part of the Deep Learning team supervised by Suvrat Bhooshan

I am working on pipelines for large-scale data-extraction, training and deployment of Text-To-Speech (TTS) models, and speech enhancers.

DAAD WISE Fellow
co-supervised by David Kappel and Anand Subramoney (and Mark Schöne!)

We extended the group's prior work [Event-SSM] to the Mamba SSM and tested it on Event-stream and Point-cloud datasets. I also helped develop a hardware-aware implementation by modifying the CUDA kernels from the original Mamba implementation. The method scales to extremely long sequences (264!) of events and is amenable to recent language-like unsupervised pretraining methods. Our work is on arXiv, I also presented this poster at IndoML'24.

Undergraduate Research Assistant
co-supervised by Ashwin Srinivasan, Tanmay Verlekar and Sidong Liu

I worked on several projects around Machine Learning for Healthcare during my time at the lab. Recently, I helped develop an implementation for the PXP protocol (also developed by the lab). We use LLMs to simulate human-LLM interactions, verify the theory and provide insights into the relation between intelligibilty performance on simple tasks in radiology and retrosynthesis [arXiv / GitHub].

We also developed a spatiotemporal graph convolution network for ataxic gait detection, which works end-to-end from raw videos to detection and severity prediction [arXiv].

We are in the process of releasing an empirical study of harmful / beneficial machine learning with real-world radiology data, inspired by the human-window hypothesis. I won the first prize in an ACM Goa event for this work's presentation [slides].

Finally, I also helped benchmark LLMs for radiology report generation and as a judge for evaluation, we also proposed a simple solution from our observations, we'll release this to arXiv soon!

Research Intern
supervised by Bharath Ramsundar

I contributed to the DeepChem library, co-developed flexible MolGAN and Normalizing Flow implementations in PyTorch [arXiv / NeurIPS Poster] and tutorials [1] [2] [3] on materials science and equivariance. I also developed a preliminary pipeline for arbitrary property optimization for materials before I left the team in January'25.

President
Previously General Secretary for the year 2023-24.

I've helped create two induction assignments for the group, [2024] and [2025]. I present frequently at the group's weekly meetings (e.g. on Deep SSMs [video]), I started a blog-posts initiative, our first blog-post on Graphormer was accepted to the GRaM Workshop at ICML'24. I have also co-led 2 AI Symposiums in [2023] and [2024]. Lastly, I helped write simple GPU runners for our reproduction of Teaching CLIP to Count to Ten.


Acknowledgement

I am immensely grateful to all of my supervisors for their guidance and support. They have played a massive role in the way I think and approach problems. In no particular order, they are Ashwin Srinivasan (Sr. Professor, BITS Goa), Tanmay Verlekar (Assistant Professor, BITS Goa), Sidong Liu (Full Professor, Macquarie University), Bharath Ramsundar (Founder / CEO, Deep Forest Sciences), David Kappel (PostDoc / Group Leader, RUB Junior Professor, University of Bielefeld), Anand Subramoney (Assistant Professor, Royal Holloway, University of London), Mark Schöne (PhD Student, TU Dresden), and all of my fellow SAiDL members & seniors!

Extra Stuff This is a list of random things, papers, tutorials, pictures, etc. that I find interesting.