Skip to Main content Skip to Navigation
Theses

Vers un apprentissage sans exemple plus réaliste

Abstract : This thesis focuses on zero-shot visual recognition, which aims to recognize images from unseen categories, i.e. categories not seen by the model during training. After categorizing existing methods into three main families, we argue that ranking methods habitually make several detrimental implicit assumptions. We propose to adapt the usual formulation of the hinge rank loss so that such methods may take inter and intra-class relations into account. We also propose a simple process to address the gap between accuracies on seen and unseen classes, from which these methods frequently suffer in a generalized zero-shot learning setting. In our experimental evaluation, the combination of these contributions enables our proposed model to equal or surpass the performance of generative methods, while being arguably less restrictive. In a second part, we focus on the semantic representations used in a large-scale zero-shot learning setting. In this setting, semantic information customarily comes from word embeddings of the class names. We argue that usual embeddings suffer from a lack of visual content in training corpora. We thus propose new visually oriented text corpora as well as a method to adapt word embedding models to these corpora. We further propose to complete unsupervised representations with short descriptions in natural language, whose generation requires minimal effort when compared to extensive attributes.
Complete list of metadata

https://tel.archives-ouvertes.fr/tel-03153445
Contributor : Abes Star :  Contact
Submitted on : Friday, February 26, 2021 - 12:58:09 PM
Last modification on : Sunday, February 28, 2021 - 3:09:42 AM

File

96284_LE_CACHEUX_2020_archivag...
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03153445, version 1

Collections

Citation

Yannick Le Cacheux. Vers un apprentissage sans exemple plus réaliste. Intelligence artificielle [cs.AI]. Conservatoire national des arts et metiers - CNAM, 2020. Français. ⟨NNT : 2020CNAM1282⟩. ⟨tel-03153445⟩

Share

Metrics

Record views

99

Files downloads

38