Min-Max Zero-Shot Multi-Label Classification

Sima BehpourICLR 2023 · OpenReview

Abstract

In many classification problems, acquiring labeled examples for many classes is difficult, resulting in high interest in zero-shot learning frameworks. Zero-shot learning (ZSL) is a problem setup, where at test time, a learner observes samples from classes that were not observed/trained in the training phase and is required to predict the category they belong to. Zero-shot learning transfers knowledge from seen classes (observed classes in the training phase) to unseen classes (unobserved classes in the training phase but present in the testing phase), reducing human labor of data annotation to build new classifiers. However, most zero-shot learning researches target single-label classification (multi-class setting). There are few studies on multi-label zero-shot learning due to the difficulty in modeling complex semantics conveyed by a set of labels. We propose a novel probabilistic model that incorporates more general feature representation (e.g., Word-Net hierarchy, word2vec features, convolutional neural network features (layer-wise), and co-occurrence statistics) and learns the knowledge transfer in terms of data structure and relations. We also investigate the effect of leveraging different CNN layers' features. Our experimental results prove the efficacy of our model in handling unseen labels. We run additional experiments to analyze the flat-sharp minima convergence of methods as a generalization factor. Our study suggests that our proposed method converges to flat minima resulting in strong generalization.
BibTeX
@inproceedings{behpour2023min,
  author = {Sima Behpour},
  title = {{Min-Max Zero-Shot Multi-Label Classification}},
  booktitle = {International Conference on Learning Representations},
  year = {2023},
  url = {https://openreview.net/forum?id=gxq1n1f0c7l},
}

Loading related content...

Figure