I've given each paper a one-sentence summary so you can see what I've been up to. You might also be interested in my Google Scholar page.
My blog
A deep dive into o3's stellar performance on the ARC-AGI benchmark, what makes it hard for LLMs, and what benchmark performance is really measuring.
Nature Scientific Reports
Exploring two new approaches to the then-stalled ARC-AGI benchmark, based on neurosymbolic reasoning and LLMs.
European Heart Journal
A new more accurate risk calculator for patients with venous thromboembolism, from analysis of UK patient data. In collaboration with ForecomAI, Pfizer, BMS and others.
European Heart Journal
A new more accurate risk calculator for patients with venous thromboembolism, from analysis of UK patient data. In collaboration with ForecomAI, Pfizer, BMS and others. (yeah there are two separate papers).
arXiv preprint
An empirical analysis of the performance ofElo, Glicko2 and TrueSkill skill rating systems in Counter-Strike: Global Offensive, based on surrogate modelling of professional matches.
Nature Communications Biology
Developed ensemble deep learning approaches for protein localization in cellular imaging, achieving state-of-the-art accuracy.
Computer Vision and Pattern Recognition (CVPR) 2023
Introduced Model Architecture Backdoors, where an exploitable deficiency can be hidden directly in the model architecture and persist through retraining.
arXiv preprint
Algorithms for disambiguating the reading order of characters in ancient Japanese manuscripts.
International Conference on Computational Creativity 2020
An annotated dataset of facial expressions in pre-modern Japanese art, suitable for classification and generative models.
Proceedings of IPSJ SIG Computers and the Humanities Symposium
The results, outcomes and learnings from the Kaggle Kuzushiji Recognition competition, which I hosted. (Paper in Japanese)
International Conference on Computer Vision, 3rd Youtube-8M Workshop
Learning to annotate events in YouTube videos, combining XGBoost and deep learning, based on weakly-labelled data.
IEEE Transactions on Circuits and Systems for Video Technology
New deep learning architectures and ensemble methods which achieved state-of-the-art results on the Youtube-8M dataset.
NeurIPS Workshop on Machine Learning for Creativity and Design
Introducing the Kuzushiji-MNIST dataset, a standardised test of visual models that overcomes the limitations of the MNIST dataset.> 700 citations.
CVPR Youtube-8M Workshop
Techniques and principles for ensembling deep learning models, based on the Youtube-8M dataset and competition.