Supervisor(s): Jen Jen Chung,Brendan Tidd,Yifei Chen

Designing optimal reward functions for RL is a challenge, especially for complex tasks like walking. Recently, LMMs have gained attention for their remarkable capability to translate natural language into machine-level instructions. ViTLearn explores the use of large multi-modal models (LMMs) to automatically generate rewards for robotic tasks.

Supervisor(s): Mashhuda Glencross,Jason Weigel,Ben Rose

SwingTheory is a tennis coaching app for beginners that analyzes users' swing posture using OpenPose and machine learning, providing corrective feedback and tracking improvement through live sections. It helps users refine their technique by comparing their form to an ideal posture.

Supervisor(s): Alina Bialkowski

Deep neural networks demand substantial computational resources and are often seen as black boxes in decision-making. Model compression and explainability can address these issues. This project investigated the interplay between pruning, accuracy and explainability, and developed an explainability-based pruning approach that revealed key differences between human and machine vision processing.

Supervisor(s): Matt D'Souza

This project creates a graph-based neural network framework written in the Rust for Machine Learning and AI applications.