Analyzing and Improving Representations with the Soft Nearest Neighbor Loss
This work addresses representation learning for machine learning practitioners by providing a tool to enhance model robustness and interpretability, though it is incremental as it expands an existing loss function.
The paper tackles the problem of analyzing and improving class manifold entanglement in representation space using the Soft Nearest Neighbor Loss, finding that maximizing entanglement in hidden layers improves generalization and uncertainty calibration, with outlier data recognized by having fewer neighbors from the predicted class.
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.