About the seminar

Flow-matching is now the ubiquitous formalism to train generative models: it's simple and effective, in addition to being very successful. But the key advantage of such formulation is for generating samples conditionally to an auxiliary information (like generating an image from its textual description) since the denoising framework avoids the pitfall of having to transform the auxiliary input into the target sample, and instead consists in driving the denoising process with the auxiliary input. This raises the question of how effective this driving is. In this talk, I will present 2 recent contributions on the subject. The first one, MIRO, shows that extra-conditions can act as specifiers for the target distribution and that having them as additional inputs speeds-up the learning. The second one, RP-FLOW, explores the transductive case and shows that examples can be leveraged as implicit extra-conditions through conditional resampling of the initial noise distribution.

Contrastive learning provides a simple and powerful framework for learning representations by defining which samples should be similar and which should be different. But what happens when these relationships are more nuanced than a binary notion of positive and negative pairs?

In this talk, I will present a line of work investigating how the structure of contrastive objectives shapes the representations they learn. I will first focus on weighted contrastive learning, showing that the weights assigned to pairwise relationships can be understood geometrically: a weighted InfoNCE objective implicitly specifies a target geometry for the representation space. This perspective connects contrastive learning to the Distance Geometry Problem and provides a framework for understanding which representation geometries can be induced by a given contrastive objective.

I will then illustrate how this perspective can be used in brain MRI representation learning. In this setting, large-scale neuroimaging datasets provide rich auxiliary information—such as anatomical measurements and phenotypic variables—that can be used during pretraining without requiring explicit labels for downstream tasks. I will show how relational information can be incorporated into a contrastive objective to shape the geometry of a brain MRI foundation model, providing a way to move beyond purely image-based self-supervision.

Finally, I will extend the discussion from weighted relationships between samples to relationships across modalities. Multimodal data contain information that may be redundant across modalities, unique to individual modalities, or only accessible through their interaction. I will present CoMM, a multimodal contrastive learning framework that explicitly accounts for these different types of information.

Together, these works suggest a broader view of contrastive learning: the contrastive objective is not merely a mechanism for pulling positives together and pushing negatives apart, but a tool for specifying the geometry and information structure that a representation should preserve.

TBA