Jamie Stirling

Postgraduate Student in the Department of Computer Science

What do you do?

I am a PhD student. My interests include compositional, few-shot and domain generalisation in deep learning, with a focus on generative models. I enjoy a mixture of theory and practical research.

How are you involved in this area of science? 

My earlier work concerns domain generalisation (DG) in deep learning, i.e. maintaining model performance on new data points that lie outside the distribution of the training data. My work focused on bias and scaling laws for DG image classification. Following this I worked on generative discrete representation learning for images, with a focus on learning space-efficient and robust image representations that are useful for downstream applications.

More recently I've worked on compositional generalisation for conditional text and image generation (e.g. text-to-image), applying probability theory to compose the outputs of discrete generative models, in order to attain more fine-grained control over generated outputs. My most recent project involves few-shot visual reasoning with very limited data, incorporating ideas from meta-learning and program synthesis.

What do you love about this topic?

I love that there's a significant amount of theory involved, but at the same time, ideas can be implemented and tested in a relatively short space of time. Working with generative models in particular is satisfying because the model outputs (such as text and image) are human-readable, so if an idea works well it will be immediately apparent. If an idea doesn't work well, it's interesting to inspect the outputs directly and see what kind of mistakes the model makes. I also like that sometimes, relatively simple ideas can make a big difference to model performance. I like the idea of only introducing complexity to a solution when absolutely necessary.

How does this work deliver real-world impact?

Improving the various kinds of generalisation in deep learning (domain, few-shot, compositional etc.) has the potential to reduce the amount of data and computational resources needed to create deep-learning based solutions. There are many real-world areas which could greatly benefit from deep learning, but data availability and compute budgets are limited. Better generalisation is the key to unlocking the potential of deep learning in these areas, because it allows us to make better use of what's already there.

Find out more:

• Read more of Jamie Stirling's work into computer vision and pattern recognition.