Background
Before starting at compass, I graduated from Bristol university with a First Class honours Masters of engineering degree in Mathematics and Computer Science. In my bachelors dissertation I did a project on simulating visual hallucination patterns in a spiking neural network in order to validate a field equation model.
For my masters dissertation I worked on meta-learning and in particular with the Model-Agnostic-Meta-Learning (MAML) algorithm. I adapted the algorithm to learn from temporal data, such that given a time series it could learn from data at various points along the time series and then construct models for a given time point with k-shot learning. I also created a novel extension to the MAML algorithm which had increased performance in the temporal setting.
Recently I completed my mini-project for the CDT which involved investigating the degree of catastrophic forgetting observed when training different network architectures in a continual learning setup. Continual learning is where you try and train a neural network on a series of tasks in a sequential manner, after finishing training on one task and training begins on the next we see that performance on the previous task plummets which is called catastrophic forgetting. Specifically we looked at the effect of freezing individual layers and the type of pretraining (trained on reconstruction task vs. classification task) on forgetting.
Research Interests
With my supervisor Dr. Rihuan Ke I am currently working on methods to train neural networks on data in a continual fashion. More specifically we are looking at methods for learning continually from unlabelled data. The dream is to create methods able to learn good representations from a continual stream of (unlabelled) data seeing each sample only once. Imagine a chatbot that can be updated to talk to you about current affairs on the fly or a robot able to continually build representations of its (constantly changing) surroundings using input from its sensors.
Email me if you want to get in touch and you can find my github here: https://github.com/h-aze