This is super cool. Nowadays experience is valued over abilities to do things. People often assume more experiences bring more abilities. But again and again many great machine learning projects are done by young team who have limited experience.

Agreed! The best Oxford entrepreneurs I know can do it without the Uni support. What they don’t realize is that to invest in the best startups, it is the University that should be begging to be part of the deal. Not the other way around.

My summer intern told me today she missed working with us. Here, she was encouraged to think of the why for everything she did daily. Now she is back in engineering, where over-optimization for the wrong thing is the norm. She spotted the difference immediately. It made my day.

Speaking of first hand experience, I know several accounts of my peers who didn't receive timely feedback until the day of the examinations. The core issue the article exposes is how Oxbridge faculties are evaluated. Teaching is simply a tickbox exercise in promotion review.

Synthetic data generation can potentially generate unlimited labelled data for deep learning systems. One of the best sources of synthetic data is the one informed by the laws of physics. Conveniently, biomechanics models can generate massive pre-training data for wearables.

#Tokenization is undeniably a key player in the success story of #LLMs but we poorly understand why. I want to highlight progress we made in understanding the role of tokenization, developing the core incidents and mitigating its problems. 🧵👇

As happened in discussions with colleagues, we don’t really have good tokenizers for sensor data as the core ml folks are busy with text and audio. But we will catch up eventually.

ML is a fast moving field. A consequence of this is that many junior ML researchers don’t receive good feedback from senior researchers who never played with the ML methods themselves to develop good intuition. Does this only happen in Oxford or across the board?

Thought experiment on a commercial biobank: 10% of Chinese population, omics, self-report, imaging and physical measurements. Selected populations will be followed up longitudinally. Participants get paid in cash and shares. Open for all use but insurance and the military.

This also applies to large orgs in general including academia and the government when evals for performance at scale are very hard.

🚩Come join our group working on new Wellcome funded project Oxford Population Health (OxPop) Big Data Institute on clinically important but neglected condition: alcohol-related dementia my.corehr.com/pls/uoxrecruit… Informal chats welcome - get in touch!

Happy to see our foundation model work for wearables recognised here but stayed tuned for our upcoming release later this year 😉

Excited to be attending #ICLR2025 next week in Singapore! Keen to connect if you are working: * Machine learning for wearables * Generalist medical AI * Multi-modal learning with large-scale datasets (e.g., biobanks) Feel free to reach out 😉

The tide has turned indeed. A few years back, most of my brightest friends are staying in the US or wanting to move over. The new generation of fresh Europe-educated PhDs are choosing to stay. That's great news for Europe.

🚨 New preprint on arXiv from OxWearables PIXL @ Oxford ! Can vision-language models (VLMs) help automatically annotate physical activity in large real-world wearable datasets (⌚️+📷, 🇬🇧 + 🇨🇳). 📄 arxiv.org/abs/2505.03374 🧵1/7

We pay too much attention to growth hacking but overlook the value of organic growth both in life and at work. Check out my latest essay on this topic: hangyuan.xyz/2025/05/19/bei…

Increasingly, text from everyone looks posh and GPT-like. Am I the only one who starts to miss the messy and imperfect text that we used to receive as that's how we are as humans?

If you are working on AI for health×timeseries please consider submitting to our NeurIPS 2025 workshop below. The team has prepared a great lineup of speakers and topics — stay tuned for updates!