Contrastive Learning with 3D Shapes

Andrea Bernini

Keywords: Deep Learning, Contrastive Learning

TL;DR: Implementation of a contrastive learning regime for a non-Euclidean data type.

Abstract:

In fields such as Computer Vision or NLP, there is a large amount of data available which, however, cannot be labeled, as it would be very expensive. A possible solution to this problem is Contrastive Learning, a Self-Supervised technique.
This work aims to implement a contrastive learning regime for a non-Euclidean data type, more precisely 3D Point Cloud Shape.