Bayesian Priors for Physically Motivated Label Smoothing in Segmentation Problems

Jamie Bernardi

Jamie studied Physics at the University of Cambridge. He did a project in deep learning and was a machine learning engineer for 2 years. He's now co-founding a non-profit and doing product development for the AGI safety fundamentals course.

Author’s Note

What was your thesis topic?

My thesis topic was in bayesian label smoothing for image segmentation, and had the aim of making medical diagnoses more accurate and interpretable.

What do you think the stronger and weaker parts of your research are?

I was quite new to machine learning, so I didn’t make any great strides in ML research. Nonetheless, I learned a lot and it helped me make my next move in the field. 

I learned a whole lot more about AI Alignment since finishing the project! I knew at the time that the project was not a particularly important contribution to alignment research, but if I was to have another go I’d be much better informed about what to contribute to. It’s also true that there are many more leverageable research topics available in 2023 than in 2018, when I started my project.

What recommendations would you make to others interested in taking a similar direction with their research?

Before deciding your thesis topic, I’d recommend learning machine learning and reading about AI alignment (perhaps by taking the AGI safety fundamentals course – I’m biased, of course!). Getting a stronger sense of the open problems in the field would have helped me a lot!

Published 5/02/23