Hierarchical Classification by Training to Diffuse on the Manifold
Keywords: hierarchical classification, diffusion, hierarchical contrastive learning
Abstract: Hierarchical classification, the problem of requiring classifying images according to a hierarchical taxonomy, has broad applications owing to the principle of ``making better mistakes'', i.e., better to predict correct coarse labels than incorrect fine labels.
Despite the importance, the literature has found it sufficient to use the wide-adopted top-1 classification accuracy to rank methods and unnecessary to use hierarchical metrics.
Importantly, the method of training deep neural networks using CE loss (or a flat softmax classifier on leaf classes only) performs well, and other ad-hoc methods do not necessarily rival the flat-softmax method.
As a result, hierarchical classification has been under-explored and there lacks training and inference methods for this problem.
In this paper, we study hierarchical classification from a novel perspective of hierarchical manifolds, assuming data from a hierarchical taxonomy lie in a hierarchical manifold in the feature space.
This motivates our novel strategies for training models of hierarchical clasification and inference.
For training, we propose a hierarchical cross-modal contrastive learning method to finetune a vision-language pre-trained model, aiming to craft an embedding space that mirrors hierarchical taxonomy.
For inference, we present a diffusion method, re-conceptualizing hierarchical classification by treating its structure as a graph.
Our method distinguishes from the conventional top-down and bottom-up methods for hierarchical classification.
Extensive experiments validate our methods, achieving the state-of-the-art on two large-scale datasets, iNaturalist2018 and iNaturalist2021, with respect to both the top-1 accuracy and diverse hierarchical metrics.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4856
Loading