Posts by Collection

I did not accept this award, but it was awarded for my work in the RNALab under Dr. Artem Babaian on using embedding models in the field of computational virology.

Turing Machine Simulator

A C++ Program providing the functionality to design, customise and emulate a single-tape Turing Machine.

Messaging Server

A small social network written in C to run on a linux server

Parallel Make

An implementation of the make program but parallelized

Breakout

A custom version of the classic Breakout game built from scratch in MIPS assembly

Flipwise

A program for planning, tracking, and splitting the cost of shared expenses

Skilltree

A web app that takes what you want to learn and constructs a personalised skill tree, dividing this goal into sub-goals - giving you milestones to achieve on your learning path.

Book Depo

A web app that aggregates free ebooks available online for download into a singular search engine

ToDo+

A desktop todo application combining multiple productivity tools into one location

Time Bandit

A 2D side-scrolling hack n’ slash game built in Godot

Script Writer

A program to assist in writing documents in the format of a play script

Part-Of-Speech Tagging

A Hidden Markov Model POS tagger with a 93.5% accuracy in predicting ground-truth tags.

Mini-C Compiler

A C-subset compiler written in C++ using ANTLR and LLVM

A Preemptive User-Level Threads Package

A custom user-level threading package in C with the ability to create, destroy, concurrently run, and preempt threads

Virtual Memory Paging Simulation

Simulating the operation of page tables and multiple page replacement algorithms.

FUSE Very Simple File System (OSTEP)

A user-space implementation of the VSFS (Very Simple File System) from OSTEP.

Battleship Solitaire Solver

A program using ARC-3 and backtracking search to solve the battleship solitaire constraint satisfaction problem

Sales and Customer Analytics Tool

A tool constructed for RJ Corp India’s Nike and Rookie division for my IB Computer Science Internal Assessment project

Mini Haskell Type Inference

Polymorphic Type Inference and Unification for a subset of the Haskell language

Paper Title Number 1

Published in Journal 1, 2009

This paper is about the number 1. The number 2 is left for future work.

Recommended citation: Your Name, You. (2009). "Paper Title Number 1." Journal 1. 1(1).
Download Paper | Download Slides

Reload: Data-Free Machine Unlearning and Model Editing through Selective Data Forgetting and Selective Data Replacement (Ongoing)

A self-driven research project undertaken as a research intern at the Vector Institute under Dr. Rahul Krishnan

Automatic labelling of data clusters: Alignment checking and unsupervised clustering

Using word embeddings to learn a descriptive feature space to automatically annotate unlabelled clusters

Understanding the dynamics of adversarial training through generative adversarial kolmogorov-arnold networks

Using the Kolmogorov-Arnold network capacity for symbolic equation decomposition to understand the phenomena of adversarial training

Are video-games effective tools for early-years mathematical education

A project completed under Prof. Steve Engels, designing a hyper casual mathematics game for reinforcing arithmetic skills in children

Exploring Syntactic Dependency Trees as a Method for Assessing Semantic Equivalence

Discovering whether there are more efficient implementations by first researching whether similarity measures on dependency trees are a good measure of semantic equivalence between sentences.

An Investigation and Simulation of the Impact of COVID-19 on the Political Satisfaction of the General Public

A sentiment analysis based project aimed at understanding political satisfaction during COVID-19 through scraping and studying reddit data

To what extent is Proximal Policy Optimisation a more efficient and accurate reinforcement learning algorithm in procedurally generated environments?

An evaluation of Proximal Policy Optimisation on procedurally generated video game environments

You should think about that some more: Recurrent Threshold Neural Networks

Exploring using a probability threshold and a feedback loop to create a dynamically deep neural network with the ability to say ‘I don’t know’

Downstream Feature Masking for Contrastive Pretrained Models (Ongoing)

A research project in collaboration with Patrik Reizinger on unlearning in contrastive self-supervised models

Are Sequential Models Fantastic Simulators? (Ongoing)

Exploring the use of autoregressive models in simulations

Can Self-Supervised Models unlearn themselves? (Ongoing)

A study into whether the data-augmentation and completion mechanisms of self-supervised models allow them to unlearn data themselves

Discovering hidden correlations from Sequence Read Archive Bioproject metadata

Conducted during my time at the RNALab under Dr. Artem Babaian. Studying the metadata in the sequence read archive using text embeddings and unsupervised learning.

Continuous and Real-Time Audio Diffusion through freeze maps (Ongoing)

Exploring the use of freeze maps to use stable diffusion to generate audio in real time

Fine-Tuning LLMs on small datasets for real-time script generation in the performing arts

Studied ways to use QLoRa fine-tuning for LLMs with small theatrical datasets

Manipulating the historical latents of large diffusion models for targeted generation (VoiceScroll Project)

Discovering how to manipulate old stored latents for generation of images consistent with the current latents

Interesting State Learning as a paradigm for Reinforcement Learning in Mario

Using a learned probability distribution to select interesting states for reinforcement learning trajectories to start from

Machine Unlearning: An Introduction

Published: August 02, 2024

This talk gave an introduction to machine unlearning for the purposes of removing biases, confusion, and protecting user privacy. It focused on the premise of unlearning, the difference between exact and approximate unlearning, and methods which approach unlearning. Additionally, I discussed major open problems in unlearning and the challenges in the field. These focused on problems regarding data, metrics and evaluation, and the need for a unified ‘goal’ for unlearning. I ended the talk by doing a deep dive into how these problems are exacerbated in the cases of Large Language Models, and recent efforts (TOFU and WMDP) to improve the quality of work in the field.

Aviraj Newatia

Posts by Collection

education

experience

grants

projects

publications

research

talks

teaching