Generalized zero-shot object recognition withoutclass-attribute relations

Er, Müslüm
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 is zero-shot learning. Inspired by the abilities of human vi-sion, zero-shot learning aims to enable recognition of novel object categories purelybased on category-wide information, which we refer to asauxiliary class information.A more modern variant, called generalized zero-shot learning, aims to build modelsthat can accurately classify novel samples of not only zero-shot classes but also thosewith supervised training examples. Most of the recent generalized zero-shot learningapproaches rely on attribute based auxiliary class information, where the attributescharacterising each class of interest needs to be defined by an oracle. In practice,this dependency greatly reduces the practicality of zero-shot learning as it is oftendifficult to define such class-attribute relationships. To bypass this requirement, in this thesis, we propose a model that requires only class names of novel classes andimplicitly learnspseudo-attributesin an end-to-end manner purely based on a set ofcandidate pseudo-attribute word embeddings. Such word embeddings are much eas-ier to collect than class-attribute annotations, as one can easily select and utilize a setof relevant words from a pre-trained language model that provides vector-space wordembeddings. Additionally, we propose a simple contrastive loss term for improvinggeneralized zero-shot learning based on simple class-to-class name similarity scores.Our experimental results show that the proposed approach yields state-of-the-art classname based generalized zero-shot learning.
Citation Formats
M. Er, “Generalized zero-shot object recognition withoutclass-attribute relations,” M.S. - Master of Science, Middle East Technical University, 2021.