Zero-shot learning via self-organizing maps
Date
25.01.2023Metadata
Show full item recordAbstract
Collecting-labeled images from all possible classes related to the task at hand is highly impractical and may even be impossible. At this point, Zero-Shot Learning (ZSL) can enable the classification of new test classes for which there are no labeled images for training. The vast majority of existing ZSL methods aim to learn a projection from the feature space into the semantic space, where all classes are represented by a list of semantic attributes. To this end, they usually try to solve a complex optimization problem. Nevertheless, the semantic features (attributes) may not be suitable to represent the images because they are derived based on human knowledge and are, therefore, abstract. Alternatively, in this study, we introduce a novel ZSL method called SOMZSL, which has its roots in Self-Organizing Maps (SOM), a famous data visualization method. In particular, SOMZSL builds two SOMs of the same size and shape, one for the feature space and one for the attribute space, and then establishes a correspondence between them. Instead of considering a direct projection between the feature space and the attribute space, which is inherently different, SOMZSL connects them through comparable intermediate layers, i.e., SOMs. In terms of performance, SOMZSL can classify novel test classes as well or even better than existing ZSL methods without dealing with a complex optimization problem, thanks to the heuristic nature of SOM on which it is based. Finally, SOMZSL uses unlabeled test images in the construction of SOMs and can thus mitigate the domain shift problem inherent in ZSL.
Source
Neural Computing and ApplicationsVolume
35URI
https://link.springer.com/article/10.1007/s00521-023-08299-1https://hdl.handle.net/20.500.12418/14625