Review Paper
Year: 2021 | Month: November | Volume: 8 | Issue: 11 | Pages: 482-487
DOI: https://doi.org/10.52403/ijrr.20211161
Zero-Shot Learning: Teaching AI to Understand the Unknown
Deekshitha Kosaraju
Independent Researcher, Texas, USA
ABSTRACT
Zero shot learning is a method in the field of machine learning that enables AI models to categorize and predict outcomes for categories they have never encountered before without the need for labeled training data specifically for those categories. This new approach tackles an obstacle in AI by eliminating the requirement, for vast amounts of data sets that are often challenging to gather. By utilizing embeddings and transfer learning alongside attribute-based learnings empowers models to apply their existing knowledge from familiar classes to unfamiliar ones. This article delves into the workings of zero shot learning and its capability to address categories that have not been encountered before in various fields like image classification and natural language processing. We will delve into the cutting-edge methods driving this groundbreaking area of study such as Domain Stacked AutoEncoders (DaSAEs) and showcase the real-world applications of zero shot learning, across diverse domains. Moreover, we explore the drawbacks of existing zero shot learning (ZSL) methods like the challenge of domain shift. Discuss upcoming strategies designed to enhance model effectiveness and precision. In essence ZSL serves as a foundation, for developing flexible and effective artificial intelligence (AI) systems that can operate successfully in constantly changing real life scenarios.
Keywords: Zero-shot learning, semantic embeddings, machine learning, domain shift, generalization, transfer learning, unsupervised learning, attribute-based learning, deep learning.
[PDF Full Text]