International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 10 Issue: 08 | Aug 2023
p-ISSN: 2395-0072
www.irjet.net
Zero Shot Learning Mudassir Ubaid1, Mohd Tanveer Hasan2 1Student, Dept. of Computer Engineering, Zakir Hussain College of Engineering and Technology, Aligarh Muslim
University, Aligarh, India
2Student, Dept. of Computer Engineering, Zakir Hussain College of Engineering and Technology, Aligarh Muslim
University, Aligarh, India --------------------------------------------------------------------------***----------------------------------------------------------------------------Abstract - This paper provides a comprehensive overview such as color, size, and shape, which are used to reason about of zero-shot learning (ZSL), a subfield of machine learning that aims to recognize and classify new objects or concepts without prior exposure during training. ZSL utilizes semantic representations to enable models to generalize to new concepts, based on the knowledge acquired in related classes. The paper discusses the two primary types of semantic representations, attribute- based and semantic space-based methods, and examines re- cent developments in ZSL, such as generative models, novel semantic representation methods, and multi-modal ZSL. The potential applications of ZSL in a variety of fields, including natural language processing, computer vision, and robotics, are also highlighted. Finally, the paper discusses the future directions and challenges of ZSL research, including the need for large-scale datasets, improved evaluation metrics, and more robust semantic representation methods. Despite facing obstacles, recent advances in ZSL have shown promising results in enabling models to recognize and classify new objects or concepts without prior exposure to them during training.
the properties of unseen classes. Classes are mapped to a highdimensional space based on their semantic relationships, such as co-occurrence and WordNet hierarchy, using semantic spacebased methods. ZSL has been implemented successfully in a variety of fields, including natural language processing, computer vision, and robotics. ZSL has been utilized in natural language processing for tasks such as sentiment analysis and text classification, in which the model must recognize new concepts not present in the training data. ZSL has been utilized in computer vision for tasks such as image classification and object detection, in which the model must recognize new objects or attributes not present in the training data. ZSL has been utilized in robotics for tasks such as object manipulation and grasping, in which the robot must recognize new objects not present in the training data. Despite its success, ZSL still faces a number of obstacles that must be resolved. Lack of large-scale datasets with sufficient labeled data for the seen classes and a di- verse set of unseen classes is one of the greatest obstacles. Another obstacle is the need for improved methods of semantic representation that can capture the complex relationships between classes. In addition, the evaluation of ZSL models remains an unresolved issue, as traditional classification metrics may not be appropriate for evaluating the performance of ZSL models.
Key Words: Zero-shot learning, Machine learning, Deep learning, Semantic embedding, Computer Vision, Unsupervised Learning, Image Classification, Few Shot Learning, Natural Language Processing, Data Augmentation, Embeddings, Generative Models, Transfer Learning, Convolutional Neural Networks (CNN).
Recent developments in ZSL have yielded promising outcomes in addressing a number of these obstacles. For instance, generative models such as generative adversarial networks (GANs) and variational autoencoders (VAEs) have been utilized to generate synthetic data for unseen classes, thereby enhancing the performance of ZSL models. In addition, recent studies have proposed novel se- mantic representation methods, such as knowledge graph- based and graph neural network-based methods, that can capture the complexity of the relationships between classes.
1.Introduction Zero-shot learning (ZSL) is a subfield of machine learning that aims to recognize and classify new objects or concepts without prior exposure during training. Traditional supervised learning methods require large amounts of labeled data to train models, which can be costly and time-consuming. ZSL, on the other hand, provides an alternative solution to this issue by enabling models to generalize to new concepts via semantic representations. ZSL is based on the concept of transfer learning, which generalizes the knowledge acquired in related classes to new ones. This is accomplished by employing semantic representations, which capture the relationships between classes and enable the model to reason about the properties of unseen classes. There are two primary types of semantic representations: attribute-based and semantic space-based methods. Each class is described by a collection of attributes,
© 2023, IRJET
|
Impact Factor value: 8.226
In this paper, we provide a comprehensive overview of the current techniques in ZSL, including the challenges and recent developments. First, we present the fundamental concepts of ZSL and the various types of semantic representations. Then, we examine the most recent developments in ZSL, such as generative models, novel semantic representation methods, and multi-modal ZSL. Further- more, we discuss the potential applications of ZSL in a variety of fields, including natural language processing, computer vision, and robotics.
|
ISO 9001:2008 Certified Journal
|
Page 351