Reproduce “When Labels Have Structure: Improving Image Classification with Hierarchy-Aware Cross-Entropy” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
April Chan, Davide D'Ascenzo, Sebastiano Cultrera di Montesano
A Hierarchy-Aware Cross-Entropy (HACE) loss function improves image classification performance by integrating known class hierarchies directly into both predicted and target distributions. This method consistently enhances accuracy across various architectures, datasets, and training paradigms.
Key insights
- The combination of prediction aggregation and ancestral label smoothing is necessary to effectively leverage hierarchies when only leaf-level supervision is present.
- Modifying the loss function itself to be hierarchy-aware can enhance classification performance, complementing architectural improvements.
- HACE encourages models to make 'better mistakes' by predicting semantically closer incorrect classes when the true class cannot be identified.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.