A prototype-based model for set classification

Classification of sets of input vectors (e.g., images and texts) is an area of ongoing interest within both computer vision and natural language processing research. In set classification framework, a set of vectors is often modeled as linear subspaces. In this contribution, we present a nearest-prototype-based approach for learning on the Grassmann manifold, consisting of such linear subspaces. This proposed strategy learns a set of subspace prototypes that captures the distinctive characteristics of each class and a set of relevance factors, while automating the subspace dimensionality selection. These traits, along with direct access to the influence of each input vector on the model’s final decision, makes it intrinsically interpretable. We have demonstrated the computational efficiency, competitive performance, and improved explainability of our proposed classifier, compared to that of the CNN and transformer-based models, through experiments on benchmark text and image datasets.