Seminar
The event has passed

Towards Uncertainty-aware Model Reduction with Martin Trapp, KTH - CANCELLED

CANCELLED due to unforeseen circumstances affecting our speaker.

As machine learning models grow in size and complexity, reducing their computational and memory requirements becomes increasingly important. But what do we lose when we make models smaller or more efficient—and how can we quantify the uncertainty this introduces? Martin Trapp, a WASP assistant professor of probabilistic machine learning at KTH Royal Institute of Technology will talk about his research. 

Overview

The event has passed
26 Oct 2026, 14:0026 Oct 2026, 15:00
Seats available:30
Language:English

CANCELLED due to unforeseen circumstances affecting our speaker.

Abstract:

The increasing complexity and number of internal states in modern machine learning models, such as KV caches in large language models (LLMs), have led to growing interest in model reduction techniques. These include quantisation, pruning, distillation, and KV cache eviction or summarisation.

While traditional benchmarking of machine learning models and model reduction techniques has often focused solely on predictive performance, I argue that uncertainties associated with model reduction should also be considered. These uncertainties are often correlated with the expected degradation in model performance.

In this talk, I will present work on variational inference in low-precision regimes and discuss how this approach can be used to quantify uncertainties arising from quantisation.

Finally, I will showcase ongoing efforts to measure the impact of KV cache eviction and the uncertainties it introduces.

Bio:

Martin Trapp is a WASP assistant professor of probabilistic machine learning at KTH Royal Institute of Technology and an incoming ELLIS Scholar at the ELLIS unit Sweden. Previously, he spent five years at Aalto University as a postdoctoral researcher and an Academy of Finland postdoctoral fellow, and completed his PhD at Graz University of Technology in Austria. His research focuses on tractable probabilistic models, such as probabilistic circuits for exact probabilistic reasoning, and on uncertainty quantification in deep learning through Bayesian deep learning and approximate inference. He is also a co-creator of Turing.jl, a probabilistic programming language in Julia.

James Bailie

  • Postdoc, Data Science and AI, Computer Science and Engineering
To personal page

Chenxiao Ma

  • Doctoral Student, Data Science and AI, Computer Science and Engineering
To personal page