profile photo

Karan Bania

 -  Updates  -  Experience  -  Projects  -  Publications  -  Contact  -  Phone  - 

I am an undergraduate student at BITS Pilani, Goa Campus where I'm pursuing Bachelor of Engineering in Computer Science. I am currently working with Prof. Ashwin Srinivasan , Prof. Sidong Liu and Prof. Tanmay Verlekar at APPCAIR on developing a generalist AI for grounded radiology reports.

I have also recently started contributing to DeepChem, currently focusing on the Materials Science part of the librtary. Thanks to my small RA at the Graph and Geometric Learning Lab at Yale, I have read several papers on GNNs focusing on vario(oo)us aspects.

I am also the General Secretary of SAiDL, Society for Artificial Intelligence and Deep Learning at BITS Goa; we try to inculcate a spirit of AI and DL in the university through open source projects and personalised courses. My research interests, clearly, revolve around Graph Neural Networks, Multimodal Multitask Learning, Machine Perception and (some) Reinforcement Learning.

In my free time you'll probably find me reading up on theoretical computer science (sorry, not sorry). I am a pretty radical supporter of R&B and rap. Apart from this, I love to teach (just about anything). I (would) also like (to) play Table Tennis, though I have not played competitively since 2019.

~ Specialist -> Generalist -> Specialising Generalist ~

 /  Email  /  CV  /  GitHub  /  LinkedIn  / 


Our Spring Induction assignment for the 2024-25 cycle is out! [Link]
Fill out this form and join the Slack workspace for any doubts. Looking forward to your submissions!


Selected for the DAAD WISE program (one of 150 undergraduates from India) at the Ruhr-Universität Bochum, Germany for summer 2024! Institut Für Neuroinformatik will be hosting me.


Selected for the MITACS GRI program (over 30,000 applicants) at the Ontario Tech University, Canada for summer 2024! VC Lab will be hosting me.


TAing the BITS F464 - Machine Learning course at BITS Goa this semester. (You can check out our labs here)


Actively maintaining LaTeX notes of our Computer Networks course you can find them here (with a Disclaimer!).


Made the General Secretary of SAiDL for the year 2023-24!


Started contributing to DeepChem.


Check out our annual AI Symposium! Register here! [linktree]


Selected as a Research Assistant at the Graph and Geometric Learning Lab at Yale university!


One of the 8 people selected for a Reading Course on LLMs instructed by Prof. Lovekesh Vig from TCS Research!


Will be an instructor for the CTE course "Intro to ML/DL" with Yash and Tejas, this semester.


Started working at APPCAIR as a student researcher!


TAing the CS F214 - Logic in Computer Science course at BITS Goa this semester.


One of four people from our batch to get inducted into SAiDL!


Selected for IIITH's 7th Summer School on AI with focus on Computer Vision.


Will be working as a ML intern at VoiceQube for summer 2023.


Selected for Neuromatch Summer School'23 in the Deep Learning Track.

Research Assistant | Graph and Geometric Learning Lab, Yale
October '23 - January '24

Presented papers to the lab and read up on a large aspects of GNNs, starting from core GNN architecture research to efficient sampling methods for few and one shot learning techniques. Also ran baselines like TransE, QuatE, etc. on KG level tasks. You can see the presentations/papers here! (Indexed according to the appropriate week/presentation number)

Undergraduate Researcher | APPCAIR & Macquarie University (Australia)
August '23 - Present

Building a generalist (multi-modal) AI for Chest X-rays (using the Stanford CheXpert dataset), which can tutor, generate and correct radiology reports; as well as retrieve images according to a radiology report. Our main focus is on Coherent Report Generation.

Jan '24 - Present

Also working on Spatio-Temporal GNNs for detecting Ataxia patients; Our main focus here si to utilise the underlying graph structure of the human body to detect anomalies in the movement of the body.

ML Intern | VoiceQube
June '23 - August '23

Worked on an AI research project to incorporate explicit sentiment analysis into effective stock market trading solutions.

Open-Source Molecular Processing Pipeline for Generating Molecules
S. Vinaya, J. Siguenza, K. Bania*, B. Ramsundar

Under review at the GEM Workshop, a part of ICLR 2024.

Image Segmentation on arbitrary prompts using Open AI's CLIP

I tried out various loss functions for segmentation (on PhraseCut dataset) and various model depths. Increased the inference time processing speed by training a light-weight decoder. Basically reproduced the CLIPSeg Paper. My results show that the model's performance can be improved vastly by using a better loss function. (~28% improvement in mIOU using DiceLoss vs using BCELoss for same number of iterations)

Reinforcement Learning as a Sequence Modelling Task

Read up on decision transformers and replaced them with other sequence models; LSTMs, GRUs and an Elman RNN and compared them to each other, with the primary goal of long-term credit assignment robustness (did also check for extrapolation) on the Hopper environment. (The figure is a display of the results on LSTMs - Medium (offline data) - Normal (Rewards), Left = Returns, Right = Episode Length)


Amazing paper + I read the proofs too 🫡

Punch, Kick, Duck!

This took me a while but I am ranked 9th in the world in this game's first level! xD

As of 11/01/24

My repository on cs224w's (Stanford Machine Learning with Graphs) assignment solutions is the only one with solutions to all problems (not just coding assignments)!

Craziest paper on GNNs so far!

(Ultra) - Towards Foundation Models For Knowledge Graph Reasoning


The most useful useless video I've seen.

🫡 * 100

MineDojo, This paper does all you can think of! (in MineCraft, xD)

✨Zero Knowledge Proofs✨

a (un)popular video.


Your newest wallpaper xD

(Very) Amazing paper on GNNs!

(GSAT) - Interpretable and Generalizable Graph Learning via Stochastic Attention Mechanism
My Slides

Multi-task optimization 👌🏻

(BLIP) - Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation

Google just being Google

(Gshard) - Scaling Giant Models with Conditional Computation and Automatic Sharding
My Slides

New paradigm on GNN's applicability

(CrysXPP) - An Explainable Property Predictor for Crystalline Materials

This template is a modification to Jon Barron's website. Find the source code to my website here.