Multi-class Classification without Multi-class Labels
This addresses the problem of reducing annotation costs for multi-class classification tasks, offering a viable alternative to traditional labeling, though it appears incremental as it builds on existing similarity-based learning paradigms.
The paper tackles multi-class classification without using class-specific labels by leveraging pairwise similarity annotations, proposing meta classification learning that optimizes a binary classifier for similarity prediction to learn a multi-class classifier. It demonstrates superior or comparable accuracy against state-of-the-art methods in supervised, unsupervised cross-task, and semi-supervised settings.
This work presents a new strategy for multi-class classification that requires no class-specific labels, but instead leverages pairwise similarity between examples, which is a weaker form of annotation. The proposed method, meta classification learning, optimizes a binary classifier for pairwise similarity prediction and through this process learns a multi-class classifier as a submodule. We formulate this approach, present a probabilistic graphical model for it, and derive a surprisingly simple loss function that can be used to learn neural network-based models. We then demonstrate that this same framework generalizes to the supervised, unsupervised cross-task, and semi-supervised settings. Our method is evaluated against state of the art in all three learning paradigms and shows a superior or comparable accuracy, providing evidence that learning multi-class classification without multi-class labels is a viable learning option.