Event Details

Statistics seminar in the School of Mathematical Sciences.

Title: Contrastive learning with InfoNCE loss, viewed as learning a Markov kernel

Abstract: Contrastive learning is a technique for obtaining vector representations of images,

text documents and other data types. It works by generating randomised transformations

of input data, creating diversity without destroying semantic content. The resulting augmented

data set is then used to learn a map from the input domain to the unit hypersphere, such

that proximity on the hypersphere reflects whatever notion of similarity between inputs

is implicitly encoded in the transformations. I will discuss contrastive learning with

perhaps the most popular loss function, InfoNCE, and explain existing theoretical

perspectives on its performance. I will explain an alternative perspective in which the

procedure is viewed as learning a Markov kernel and outline associated generalisation

analysis.