(paper 31) Invariance Propagation

Unsupervised Representation Learning by Invariance Propagation

Contents

1. Abstract
2. Introduction
3. Method
   1. Self-labeling

0. Abstract

Unsupervised learning based on contrastive learning

• aim to learn representations, invariant to instance-level variations

propose Invariance Propagation

• focus on learning representations, invariant to category-level variations

• recursively discovers semantically consistent samples, which are in the same high-density regions
• hard sampling

combining (1) clustering + (2) representation learning

\(\rightarrow\) doing it naively…leads to degenerate solutions