Adaptive Multi-Prototype Representation Learning

Existing representation learning usually neglects two important issues, the intra-class representation diversity and underexploited label utilization, especially the negative feedback during training process. Fortunately, prototype learning, which summarizes the samples commonality, can potentially raise label utilizations and encourage intra-class diversity. In this paper, we investigate the intra-class diversity and effective updates in prototype learning. Specifically, we propose an Adaptive Multi-Prototype Representation Learning, by incorporating the label awareness into both prototype formation and process to improve the representation quality. We separate the prototype updates from the representation optimization and exploit the label indexes to directly implement the prediction feedback. Experiments verify the effectiveness of our label-aware and joint multi-prototype updating strategies.