Abstract

Invariance is crucial for neural networks, enabling them to generalize effectively across variations of the input data by focusing on key attributes while filtering out irrelevant details. In this thesis, we study representation learning in neural networks through the lens of structured invariance. We start by studying the properties and limitations of the invariance that neural networks can learn from the data. Next, we develop a method to extract the structure of invariance learned by a neural network, providing a more nuanced analysis of the quality of learned invariance. In the next chapter, we focus on contrastive learning, demonstrating how more structured supervision results in a better quality of learned representations. The last two chapters that follow, focus on practical aspects of representation learning with structured invariance in computer vision.