- TL;DR: We propose Hierarchical Complement Objective Training, a novel training paradigm to effectively leverage category hierarchy in the labeling space on both image classification and semantic segmentation.
- Abstract: Hierarchical label structures widely exist in many machine learning tasks, ranging from those with explicit label hierarchies such as image classification to the ones that have latent label hierarchies such as semantic segmentation. Unfortunately, state-of-the-art methods often utilize cross-entropy loss which in-explicitly assumes the independence among class labels. Motivated by the fact that class members from the same hierarchy need to be similar to each others, we design a new training diagram called Hierarchical Complement Objective Training (HCOT). In HCOT, in addition to maximizing the probability of the ground truth class, we also neutralize the probabilities of rest of the classes in a hierarchical fashion, making the model take advantage of the label hierarchy explicitly. We conduct our method on both image classification and semantic segmentation. Results show that HCOT outperforms state-of-the-art models in CIFAR100, Imagenet, and PASCAL-context. Our experiments also demonstrate that HCOT can be applied on tasks with latent label hierarchies, which is a common characteristic in many machine learning tasks.
- Code: https://github.com/iclr2020-HCOT-anonymized-submission/Hierarchical-Complement-Objective-Training
- Keywords: category hierarchy, optimization, entropy, image recognition, semantic segmentation, deep learning