Representation learning for deep clustering and few-shot learning

Deep learning’s ability to extract meaningful representations from complex inputs has significantly advanced machine learning. In this talk, we will focus on representation learning in label-scarce scenarios, where the absence of a strong supervisory signal forces models to rely more on intrinsic properties of the data and its representations. In particularly, we will consider learning in the presence of few or no labels from a representation learning perspective. We will first discuss approaches to deep clustering in the multi-view setting, where naively levering contrastive alignment can significantly degrade performance when information is scattered across multiple views. Subsequently we explore how geometrical properties of representations influence few-shot classification performance. We will focus on embedding representations on the hypersphere, a geometry particularly suitable for few-shot learning, and connect this choice of geometry to the hubness phenomenon in high-dimensional spaces.

Post Talk Link:  Click Here

Passcode: %.F0tf^7