Hyperspherical Prototype Networks
This work addresses the challenge of integrating classification and regression in machine learning, offering a novel approach for multi-task problems, though it appears incremental as it builds on prototype-based methods.
The paper tackles the problem of unifying classification and regression by proposing hyperspherical prototype networks, which use predefined prototypes on hyperspheres to handle both tasks with a single loss function, enabling joint training and showing experimental benefits over existing methods.
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.