Hyperspherical Prototype Networks

TL;DR

Hyperspherical prototype networks unify classification and regression tasks by defining class prototypes on hyperspheres, enabling flexible, multi-task learning with improved performance over traditional methods.

Contribution

The paper introduces hyperspherical prototypes with data-independent positioning, allowing joint classification and regression without prototype updates or size constraints.

Findings

Improved accuracy in classification and regression tasks.

Effective multi-task learning with unified loss function.

Outperforms traditional prototype, softmax, and MSE methods.

Abstract

This paper introduces hyperspherical prototype networks, which unify classification and regression with prototypes on hyperspherical output spaces. For classification, a common approach is to define prototypes as the mean output vector over training examples per class. Here, we propose to use hyperspheres as output spaces, with class prototypes defined a priori with large margin separation. We position prototypes through data-independent optimization, with an extension to incorporate priors from class semantics. By doing so, we do not require any prototype updating, we can handle any training size, and the output dimensionality is no longer constrained to the number of classes. Furthermore, we generalize to regression, by optimizing outputs as an interpolation between two prototypes on the hypersphere. Since both tasks are now defined by the same loss function, they can be jointly trained for multi-task problems. Experimentally, we show the benefit of hyperspherical prototype networks for classification, regression, and their combination over other prototype methods, softmax cross-entropy, and mean squared error approaches.