←
Timothy Hospedales144 posts
Timothy Hospedales avatar

Timothy Hospedales

@tmh31

Professor @ University of Edinburgh. Head of Samsung AI Research Centre, Cambridge.

◷ Joined August 2009▧ 4 Photos and videos
Radamés Ajna avatar
⟳ Timothy Hospedales retweeted
Radamés Ajna@radamar·8 Dec 2023

Here's the demo "Enhance This"! It's a surreal image magnifier that creates a high-res version by imagining new details, using the SDXL base model. Thanks to @RuoyiDu's DemoFusion research. It takes a ~minute to generate a 2024x2024 image. huggingface.co/spaces/radame…

0:35
Timothy Hospedales avatar
Timothy Hospedales@tmh31·2 Oct 2023

Interested in practical uncertainty quantification? Our new Bayesian NN library from Samsung AI Cambridge scales to large VITs! One line of code wraps any architecture with no modifying your model definition! arxiv.org/abs/2309.12928 github.com/SamsungLabs/Bayes… @minyoungkim21

Raman Dutt avatar
⟳ Timothy Hospedales retweeted
Raman Dutt@RamanDutt4·16 May 2023

🚨Parameter Efficient Fine-Tuning has been well researched for NLP, vision, and cross-modal tasks. So why should MedAI be left behind? Presenting the first evaluation on PEFT for medical AI - arxiv.org/abs/2305.08252 16 PEFT methods, 5 datasets including a text-to-image task 🔥

Henry Gouk avatar
⟳ Timothy Hospedales retweeted
Henry Gouk@henrygouk·5 Jan 2023

Hello, everyone! We will be organizing an online workshop at ICLR 2023 aimed at one question: What do we need for successful domain generalization? The workshop will include invited talks from David Lopez-Paz, @AmosStorkey, @tommasi_tatiana, and @ylqzd2011 1/2

Post attachment
Timothy Hospedales avatar
Timothy Hospedales@tmh31·7 Oct 2022

A take home is that model selection strategy seems to be the only reliable way to affect fairness so far. Surprising? Good/Bad? At least it's easy to implement.