Closed-form sample probing for training generative models in zero-shot learning

Download
2022-2-10
Çetin, Samet
Generative modeling based approaches have led to significant advances in generalized zero-shot learning over the past few-years. These approaches typically aim to learn a conditional generator that synthesizes training samples of classes conditioned on class embeddings, such as attribute based class definitions. The final zero-shot learning model can then be obtained by training a supervised classification model over the real and/or synthesized training samples of seen and unseen classes, combined. Therefore, naturally, the generative model ideally needs to produce not only relevant samples, but also those that are sufficiently informative for classifier training purposes. However, existing approaches rely on approximations or heuristics to enforce the generator to produce class-specific samples. In this thesis, we propose a principled approach that shows how to directly maximize the value of training examples for zero-shot model training purposes, by inferring and evaluating the closed-form ZSL models at each generative model training step, which we call sample probing. This approach provides a way to validate the quality of generated samples in an end-to-end manner, where the generator receives feedback directly based on the prediction made on the real samples of unseen classes. Our experimental results show that sample probing improves the recognition results when integrated into state-of-the-art baselines.

Suggestions

Closed-form sample probing for learning generative models in Zero-shot Learning
Çetin, Samet; Baran, Orhun Buğra; Cinbiş, Ramazan Gökberk (2022-04-25)
Generative model based approaches have led to significant advances in zero-shot learning (ZSL) over the past few years. These approaches typically aim to learn a conditional generator that synthesizes training samples of classes conditioned on class definitions. The final zero-shot learning model is then obtained by training a supervised classification model over the real and/or synthesized training samples of seen and unseen classes, combined. Therefore, naturally, the generative model needs to produce not...
A Recurrent and Meta-learned Model of Weakly Supervised Object Localization
Sariyildiz, Mert Bulent; Sumbul, Gencer; Cinbiş, Ramazan Gökberk (2022-01-01)
© 2022 IEEE.The object localization and detection has improved greatly over the past decade, thanks to developments in deep learning based representations and localization models. However, a major bottleneck remains at the reliance on fully-supervised datasets, which can be difficult to gather in many real-world scenarios. In this work, we focus on the problem of weakly-supervised localization, where the goal is to localize instances of objects based on simple image-level class annotations. In particular, i...
Feature enhancement with deep generative models in deep bayesian active learning
Duymuş, Pınar Ezgi; Ertekin Bolelli, Şeyda; Department of Computer Engineering (2022-9)
Data-intensive models emerge as new advances in Deep Learning take place. However, access to annotated datasets with many data points is not constantly prevalent. This situation emphasizes the need for Active Learning to select the least possible amount of data without compromising the accuracy of the classifier models. Recent advancements occur in Deep Bayesian Active Learning (DBAL), which means incorporating uncertainty of model parameters into a Deep Network. In this work, we present an algorithm that i...
Out-of-Sample Generalizations for Supervised Manifold Learning for Classification
Vural, Elif (2016-03-01)
Supervised manifold learning methods for data classification map high-dimensional data samples to a lower dimensional domain in a structure-preserving way while increasing the separation between different classes. Most manifold learning methods compute the embedding only of the initially available data; however, the generalization of the embedding to novel points, i.e., the out-of-sample extension problem, becomes especially important in classification applications. In this paper, we propose a semi-supervis...
Generalized zero-shot object recognition withoutclass-attribute relations
Er, Müslüm; Cinbiş, Ramazan Gökberk; Department of Computer Engineering (2021-2-11)
Over the last decade, great improvements have been achieved in image classifica-tion performances following the advances in supervised deep learning approaches.These supervised approaches, however, typically require substantial amounts of la-beled training examples. Collecting and annotating such examples is a cumbersomeand error-prone task, especially when a large number of classes needs to be spanned.One of the promising approaches towards overcoming this limitation of supervisedrecognition techniques i...
Citation Formats
S. Çetin, “Closed-form sample probing for training generative models in zero-shot learning,” M.S. - Master of Science, Middle East Technical University, 2022.