Hello! I'm
Dr. Anna Hughes.
I'm a climate & sustainability focused machine learning engineer with a research background in astrophysics.
Machine Learning & Data Science
Neural Network Compression - Large Language Models
Large language models owe their sophistication not just to the underlying algorithms and training data, but also to their large number of trainable parameters. The rise of LLMs and other large generative models comes at a extraordinary cost; the energy required during the training and fine-tuning processes.
I led a team of 6 researchers to find solutions. As a team, we developed 4 novel approaches to neural network compression in large language models. We used machine learning and optimisation to identify and remove weak neurons from the network.
Time Series Analysis
I have worked extensively with time series data in multiple projects and across industries. Examples include:
Modeling variability in stellar flux
Modeling and forecasting with climate data
Anomaly detection in network flow log data
Anomaly detection in time series EEG data
Forecasting with financial data
Anomaly Detection
Anomaly detection algorithms - designed to identify anomalous data or events in some dataset - are immensely useful across a range of fields. I have experience identifying anomalous data using:
generative adversarial networks
isolation forests
cluster-based methods
residual neural networks
Nonlinear Regression
Nonlinear regression played a pivotal role in my Ph.D. research. I used radio observations of low-mass stars to
identify likely underlying theoretical astrophysical models
fit model parameters based on new data
make predictions about future observations and stellar behaviour
CO2 Emissions Forecasting
Quantum Computing
Quantum Machine Learning
Quantum machine learning integrates quantum components into machine learning problems; typically classical data is converted into quantum data, a set of computations is performed, and the outgoing data is converted back into classical data.
I have experience using both quantum annealers and gate-based quantum computers for clustering, support vector machines, and neural networks.
Quantum Inspired Optimisation
While quantum computers are still in their infancy, quantum-inspired algorithms such as quadratic unconstrained binary optimisation (QUBO) can be used to solve a wide range of optimisation problems with classical hardware.
I have extensive experience identifying applications of QUBO to a variety of problems such as neural network compression, feature selection, and representative selection.
Astrophysics
