Im Rahmen der Forschungsseminare der Stochastik trägt Dr. Dmitry Kobak, Universität Tübingen über das Thema: “From neighbor embeddings to contrastive embeddings” vor.

Davor gibt es ab 15:30 in Raum 200 (Helmholtzstraße 18) noch die Gelegenheit, sich zu einem Kaffee oder Tee zu treffen.

Das Thema ist für alle geeignet.

Abstract:

In this talk, I am going to present our ongoing work on neighbor embeddings and on low-dimensional contrastive learning.  In recent years, neighbor embedding methods like t-SNE and UMAP have become widely used across several application fields, from single-cell biology to deep learning engineering. Despite their success, they can only be applied when a suitable similarity metric is available. To get around this limitation, we can borrow the self-supervised learning approach from computer vision where a representation is learned based on data augmentations. In this talk I am going to explain how neighbor embeddings and contrastive learning can be unified into a single coherent framework. This will lead to our recent work on contrastive visualizations of image datasets. I will also discuss various trade-offs between neighbor embedding methods, such as the attraction-repulsion spectrum, appearing in both contrastive and non-contrastive setting. Finally, I will demonstrate some practical applications of these methods in our applied work, ranging from biology to natural language processing, and discuss future research directions and challenges.