Leveraging Class Hierarchies with Metric-Guided Prototype Learning

Vivien Sainte Fare Garnot, Loic Landrieu

Keywords: hierarchical classification, prototypical networks

Abstract: In many classification tasks, the set of classes can be organized according to a meaningful hierarchy. This structure can be used to assess the severity of confusing each pair of classes, and summarized under the form of a cost matrix which also defines a finite metric. We propose to integrate this metric in the supervision of a prototypical network in order to model the hierarchical class structure. Our method relies on jointly learning a feature-extracting network and a set of class representations, or prototypes, which incorporate the error metric into their relative arrangement in the embedding space. We show that this simultaneous training allows for consistent improvement of the severity of the network's errors with regard to the class hierarchy when compared to traditional methods and other prototype-based strategies. Furthermore, when the induced metric contains insight on the data structure, our approach improves the overall precision as well. Experiments on four different public datasets—from agricultural time series classification to depth image semantic segmentation—validate our approach.

One-sentence Summary: We introduce a new prototypes-based method to model class hierarchies in classification tasks, resulting in a decrease of both the average cost of misclassifications and the overall rate of errors.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Reviewed Version (pdf): https://openreview.net/references/pdf?id=GjMYTw4RsM