Representation Alignment in Neural Networks

TL;DR

This paper reveals that neural network representations tend to align their top singular vectors with targets after training, which explains their transferability and impacts learning efficiency.

Contribution

It uncovers the phenomenon of representation alignment in neural networks and analyzes its emergence across architectures and layers, linking it to transfer performance.

Findings

Alignment emerges across various architectures and optimizers.

Deeper layers show increased alignment.

High-performance CNNs exhibit high alignment levels.

Abstract

It is now a standard for neural network representations to be trained on large, publicly available datasets, and used for new problems. The reasons for why neural network representations have been so successful for transfer, however, are still not fully understood. In this paper we show that, after training, neural network representations align their top singular vectors to the targets. We investigate this representation alignment phenomenon in a variety of neural network architectures and find that (a) alignment emerges across a variety of different architectures and optimizers, with more alignment arising from depth (b) alignment increases for layers closer to the output and (c) existing high-performance deep CNNs exhibit high levels of alignment. We then highlight why alignment between the top singular vectors and the targets can speed up learning and show in a classic synthetic transfer problem that representation alignment correlates with positive and negative transfer to similar and dissimilar tasks.