Lunch talk by Ching Lam Choi

Title

Seeing Beyond the Cave: Asymmetric Views of Representation

 

Abstract

Multimodal learning often assumes symmetry: that representations align across modalities, that generation is reversible, and that cycle consistency holds. But when we iterate real-world generative systems—caption to image to caption to image—these assumptions break down. Instead, we observe irreversible drift, semantic collapse, and asymmetric distortions that current metrics fail to capture. 

In this talk, I propose a framework for modelling such phenomena using quasimetric flows: directional, non-conservative transitions between modalities that describe structured – often degenerative – paths in representation space. This view challenges the symmetry ideal; formalises semantic equilibrium and information decay; finally, offers a geometric lens on alignment, degeneracy and multimodal semantics.

 

Bio

Ching Lam is a PhD student at MIT CSAIL, co-supervised by Stefanie Jegelka, Phillip Isola and Antonio Torralba. Her research seeks to align multiple modalities from perspectives of symmetries, disentanglement and compositional generalisation. Prior to MIT, she studied her bachelor’s at CUHK, where she spent wonderful years researching at the Multimedia Lab (MMLab), Max Planck Institute for Intelligent Systems (MPI-IS Tübingen), MILA, Pioneer Centre for AI (Copenhagen), and others detailed on her website: https://chinglamchoi.github.io/cchoi/.

 

 

Lunch will be served at the talk. If you wish to attend, please contact Björg Birkholm Magnúsdóttir (bjma@di.ku.dk).