PGR Seminar: Berné Nortier and Mirza Hossain
You are warmly invited to the next PGR seminar:-
Wednesday 07/10/2026
14:00-15:00
JC 1.33A
Speaker 1: Berné Nortier
Distilling the Data Deluge: Network Renormalisation as a technique for coarse-graining generalised models of complex systems (Work in progress)
Abstract: Complexity surrounds us. While expanding data storage highlights just how interconnected most systems are, the sheer scale of these networks makes them increasingly difficult to study directly, highlighting the pressing need for tools that can analyse such systems at coarser levels.
Here, we develop a framework for the investigation of complex networks at different resolutions by formulating the geometric Renormalisation Group for bipartite, higher-order and metadata-rich networks. We do this by embedding them into an underlying hidden metric space and derive attributes of downscaled replicas which hold true for the original system. We then show evidence of geometric scaling under this transformation for empirical networks and discuss applications of this multi-resolution approach to large-scale systems where downscaled network replicas could serve as alternatives for analysis and for fast-track exploration of rough parameter spaces.
Bio: Berné is a 3rd-year PhD student under supervision of Prof. Simon Dobson. She is interested in structural aspects of modelling complex systems where group interactions are present, with particular focus on how latent-space and geometrical models can be used to better understand enriched network representations such as hypergraphs and bipartite networks.
Speaker 2: Mirza Hossain
AUC Is Not Enough: Compression can preserve classifier performance while altering interpretable structure in pathology embeddings
Abstract: Foundation models have become the standard way to turn pathology slides into data: each tissue patch is encoded as a high-dimensional embedding, and downstream models are trained on those frozen vectors. At cohort scale this creates a real storage and transfer bottleneck, and compression is the obvious fix. Our earlier work showed that nine dimensions are enough to recover 99% of classification AUC across several datasets and models.
But these embeddings are increasingly reused for mechanistic interpretability workflow. Sparse autoencoders (SAE) decompose embeddings into features, and pathologists review each feature by looking at the patches that activate it most strongly. This work asks whether compression that preserves AUC also preserves the evidence those experts reviewed.
Using 100 pathologist-reviewed features from a SAE trained on UNI embeddings, we compare PCA, random projection and a UMAP-based autoencoder across four tiers of preservation: geometry, gating, patch evidence and tissue-class semantics. We find an order-of-magnitude gap. PCA recovers baseline AUC with 4 dimensions but needs 47 to recover even half of the reviewed patches. At 9 dimensions, where AUC matches the uncompressed baseline, only about 10% of the original top patches survive, even though every reviewed feature is still active. The conclusion is that downstream performance alone cannot certify that an interpretation survives compression. This work was presented at the MI4MedFM workshop at MICCAI 2026.
Bio: I am a final-year PhD student in the School of Computer Science at the University of St Andrews, supervised by David Harris-Birtill. My research sits at the intersection of mechanistic interpretability and computational pathology. I focus on whether the internal features of pathology foundation models can be verified against expert judgement, and on whether that structure survives transformations such as fusion and compression.