newest/featured

State-space models can learn in-context by gradient descent

Working on using a novel SSM (State Space Model) architecture for language modeling, by using SSMs to emulate gradient descent, and exploring the mechanisms by which SSMs perform in-context learning. Under review at ICLR 2025.

Supervisors: Prof. Anand Subramoney (Royal Holloway, University of London)

RadioLM - Radiology Language Model

Working on a novel pseudo replacement to RLHF through prompting which also involves creating a modified model with this pseudo-RLHF and then using it to find a Human-window of LLM explanation understandability. We are testing this theory on Radiology/Med students.

Supervisors: Prof. Ashwin Srinivasan, Prof. Sidong Liu (Macquarie University, Australia), Prof. Tanmay Verlekar

CountCLIP - [Re] Teaching CLIP to Count to Ten

I conducted a reproducibility study of the paper Teaching CLIP to Count to Ten, published by Google Research, in ICCV 2023. I implemented the paper from scratch and collected a specialized dataset to facilitate the training. In addition to this, I carried out further explorations and analysis of the paper, and wrote a paper on my findings which is currently under review at ReScience C 2024.

[preprint] [code] [dataset]

AutoPAC - Automatic Plan and Code Synthesis

An LLM-based pipeline to apply a idea to resolve challenges in an ML pipeline. AutoPAC models a more realistic setting of incremental development of ML pipelines, resolving the issues in a continual fashion.

Supervisors: Prof. Ashwin Srinivasan, Prof. Gautam Shroff (TCS Research), Prof. Tanmay Verlekar

[preprint]

Visualising Image Generation using Stable Diffusion

I implemented the Stable Diffusion paper from scratch (with the help of this tutorial), and added the functionality to animate the image generation process. More animated generations can be found here.

[repo]

Relevant XKCD

I made a website to find the most relevant XKCD comic based on the prompt you type. Embeddings of the comic explanations were made, and the most relevant comics are fetched using a similarity search. The website is a Flask application.

[repo][link]

Variations of Softmax on a CNN Classifier

I tested variations of the Softmax activation function on a CNN classifier. I found that taking the logarithm of a variant increased accuracy and converged faster.

[repo][link]

xcode

These are much older Swift projects made from 2018 to 2020. All the featured apps were made without following a tutorial for their logic (spaghetti code warning :P). The image assets for the graphics are also made by me. I have made many more (albeit simpler) applications in Swift, the source code to which can be found below in a larger repository:
[main repo]

courses

a more exhaustive list can be found [here]