Analyzing and Improving Representations with the Soft Nearest Neighbor Loss

TL;DR

This paper introduces the Soft Nearest Neighbor Loss to measure and manipulate the entanglement of class representations, revealing that increasing entanglement in hidden layers enhances discrimination, generalization, and uncertainty calibration.

Contribution

It extends the Soft Nearest Neighbor Loss as an analytical tool and demonstrates that maximizing entanglement improves model performance and uncertainty estimation.

Findings

Maximizing entanglement improves classification accuracy.

Entangled representations lead to better outlier detection.

Enhanced calibration of uncertainty estimates.

Abstract

We explore and expand the $\textit{Soft Nearest Neighbor Loss}$ to measure the $\textit{entanglement}$ of class manifolds in representation space: i.e., how close pairs of points from the same class are relative to pairs of points from different classes. We demonstrate several use cases of the loss. As an analytical tool, it provides insights into the evolution of class similarity structures during learning. Surprisingly, we find that $\textit{maximizing}$ the entanglement of representations of different classes in the hidden layers is beneficial for discrimination in the final layer, possibly because it encourages representations to identify class-independent similarity structures. Maximizing the soft nearest neighbor loss in the hidden layers leads not only to improved generalization but also to better-calibrated estimates of uncertainty on outlier data. Data that is not from the training distribution can be recognized by observing that in the hidden layers, it has fewer than the normal number of neighbors from the predicted class.