»Ê¼Ò»ªÈË

My interest in machine learning, along with my desire to continue studying the physical sciences, particularly astrophysics, attracted me to the CDT here at »Ê¼Ò»ªÈË.

Project title: Probabilistic Deep Learning for Cosmology and Beyond

Research Group:ÌýAstrophysics

Supervisor(s):

Introduction:Ìý

I completed my MPhys at the University of Leeds in 2023. My master’s thesis, working with the molecular nanophysics group, developed a data processing pipeline and convolutional neural network for classifying images of mechanically deformed cells. My interest in machine learning, along with my desire to continue studying the physical sciences, particularly astrophysics, attracted me to the CDT here at »Ê¼Ò»ªÈË.

---

Project description:ÌýÌý

My work currently focuses on developing improved methods for Bayesian model selection, which is a statistical method used in areas such as cosmology to decide what model best describes a dataset. Bayesian model selection requires the calculation of the Bayesian evidence, which is computationally expensive, especially for high dimensional parameter-spaces. My research, working with the SciAI research group here at »Ê¼Ò»ªÈË, will leverage Monte Carlo sampling methods, deep learning, and optimisation techniques to tackle this problem.

First year group project:ÌýMediaTek Research

SwiftFish: A Unified Approach to Model Compression using FishLeg with MediaTek Research.

This project developed a neural network compression method by utilizing second-order information, provided by the FishLeg optimizer developed at MediaTek Research. With increasingly large neural network models, such as the latest LLMs, which have up to trillions of parameters, neural networks are becoming increasingly more expensive to run and store in memory. By utilizing second-order information, neural networks can be compressed to a larger degree while retaining more of their original performance.

Placement:Ìý

---

Ìý

Ìý