News

• Oct 2025
  Started research internship at AWS, working on KV cache compression for long context LLM with Matthias Seeger.
• Apr 2025
  Released the SUQ library for streamlined uncertainty quantification.
• Jan 2025
  One paper accepted at ICLR 2025.
• Oct 2024
  Three workshop papers accepted at NeurIPS 2024.

Selected Publications

See full list of publications here.

• 
  Streamlining Prediction in Bayesian Deep Learning

  Rui Li, Marcus Klasson, Arno Solin, Martin Trapp

  International Conference on Learning Representations (ICLR), 2025.

  paper / code

  While estimating posterior has been actively researched in Bayesian deep learning (BDL), how to make predictions with posterior efficiently is largely overlooked. We examine streamlining prediction in BDL through a single forward pass without sampling. We showcase our approach for both MLP and transformer models, such as ViT and GPT-2.

• 
  Post-hoc Probabilistic Vision–Language Models

  Anton Baumann, Rui Li, Marcus Klasson, Santeri Mentu, Shyamgopal Karthik, Zeynep Akata, Arno Solin, Martin Trapp

  Under Review.

  paper / code

  While Vision–Language models have shown remarkable performance in various tasks, the lack of uncertainty estimation makes them unreliable in high-stakes applications. We propose a post-hoc uncertainty quantification method based on Laplace approximation, which provides useful predictive uncertainties and better calibration.

• 
  Improving Hyperparameter Learning under Approximate Inference in Gaussian Process Models

  Rui Li, Ti John, Arno Solin

  International Conference on Machine Learning (ICML), 2023.

  paper / code

  Variational inference and expectation propagation are two commonly used approximate inferences in Gaussian process models with complementary advantages. We developed a hybrid training procedure to bring the best of both worlds.

• 
  Flatness Improves Backbone Generalisation in Few-shot Classification

  Rui Li, Martin Trapp, Marcus Klasson, Arno Solin

  Winter Conference on Applications of Computer Vision (WACV), 2025. Oral.

  paper / code

  In few-shot classification most efforts focus on adapting the backbone to the target domain without considering the importance of backbone training. We show that flatness-aware backbone optimisation can lead to better generalisation through theoretical and empirical results.

Notes / Blog

See technical notes and blogs here.