Research progress of zero-shot learning
- PDF / 645,672 Bytes
- 15 Pages / 595.276 x 790.866 pts Page_size
- 87 Downloads / 195 Views
Research progress of zero-shot learning Xiaohong Sun 1,2 & Jinan Gu 1 & Hongying Sun 2 Accepted: 7 November 2020 # Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Although there have been encouraging breakthroughs in supervised learning since the renaissance of deep learning, the recognition of large-scale object classes remains a challenge, especially when some classes have no or few training samples. In this paper, the development of ZSL is reviewed comprehensively, including the evolution, key technologies, mainstream models, current research hotspots and future research directions. First, the evolution process is introduced from the perspectives of multishot, few-shot to zero-shot learning. Second, the key techniques of ZSL are analyzed in detail in terms of three aspects: visual feature extraction, semantic representation and visual-semantic mapping. Third, some typical models are interpreted in chronological order. Finally, closely related articles from the last three years are collected to analyze the current research hotspots and list future research directions. Keywords Zero-shot learning . Feature extraction . Semantic representation . Visual-semantic mapping
1 Introduction Traditional visual recognition technologies usually require a large number of labeled training data, such as thousands of images or even tens of thousands of images for each class, to achieve good classification accuracy. However, it is impossible to collect such a large amount of data with labels for all classes, and this is a huge challenge for traditional supervised learning. At present, there are some different solutions, for example, semi-supervised learning [1], transfer learning [2], self-taught learning [3], and few-shot learning [4]. It has been found that the fewer training samples there are, the more difficult recognition is; when there are no training samples, the process is called zero-shot learning (ZSL) [5]. Human beings have the ability to retain or accumulate acquired knowledge from past learning tasks and quickly integrate this knowledge to solve new recognition tasks; this provides inspiration for researchers to attempt to endow machines with the same intelligence, for which the ability of continuous * Jinan Gu [email protected] * Hongying Sun [email protected] 1
School of Mechanical Engineering, Jiangsu University, Zhenjiang 212000, China
2
School of Mechanical Engineering, Anyang Institute of Technology, Anyang 455000, China
learning/lifelong learning [6] is needed. In this process, ZSL is of great significance for the realization of the ultimate goal of “intelligence”. ZSL can be regarded as a special type of cross-modal retrieval learning. It is a method for the recognition of new classes using unlabeled data, and the basic idea is to transfer knowledge from seen classes to unseen classes by sharing attributes. Most existing ZSL methods use an intermediate semantic representation (such as visual features or semantic word vectors), and the semantic representation is shared b
Data Loading...
