About me

I’m a machine learning engineer in Brighton, UK.

I design and build production machine learning systems end to end: from whiteboarding the problem, through training models and writing the services that run them, to automating the monitoring that keeps them improving. Recently that has meant forecasting for consumer sports products: constrained optimisation over predicted outcomes, fed by noisy third-party data against fixed deadlines, where prediction quality is the product.

The part I find most interesting is the part that isn’t modelling: the systems and processes that let a team ship models with confidence. Before a model ships, that means time-based evaluation splits and domain validation metrics, checked through backtesting. Once it’s live, it means catching it degrading quietly rather than failing loudly: accuracy slipping, or predictions collapsing onto a single value. That tends to matter more than the architecture, and gets written about less.

Research

My PhD, at the University of Sussex under Prof. Novi Quadrianto, was on fair representations that system designers and stakeholders can inspect, rather than an uninterpretable latent space, so the changes made for fairness can be interpreted and explained to the people who have to act on them. You can read the thesis here.

That thread (making machine learning legible and operable by people who aren’t the person who built it) runs through most of what I do, research or otherwise.

• Discovering Fair Representations in the Data Domain, CVPR 2019
• Null-Sampling for Interpretable and Fair Representations, ECCV 2020
• An Algorithmic Framework for Positive Action, ACM EAAMO 2021

Open source

I wrote and maintain EthicML, an open-source benchmarking framework for fair machine learning: common loaders for the standard fairness datasets, a built-in train/validation/test protocol, pluggable models and metrics, and automatic utility–fairness trade-off plots. It started because the lab was sharing code by emailing snippets, and it ended up being used in teaching and behind the experiments in published papers.

Elsewhere

📫 Reach me at oliverthomas@outlook.com.

😄 Pronouns: he/him.

⚡ Fun fact: the dress was gold and white. There was no blue and black dress.