International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 11 Issue: 05 | May 2024
www.irjet.net
p-ISSN: 2395-0072
LEARNING LIKE HUMANS: THE PROMISE OF FEW-SHOT LEARNING IN AI Arpit Shrivastava Meta Platform Inc., USA --------------------------------------------------------------------------***---------------------------------------------------------------------------ABSTRACT: Few-shot learning is a big change in artificial intelligence (AI). It shows promise for making AI more like humans in terms of being able to learn quickly and with little information. This revolutionary method goes against the usual practice of using huge datasets to train AI models. Instead, it focuses on creating methods that let models learn new things from just a few examples. This paper explores the ideas behind few-shot learning, how it works, and the different ways it could be used. It shows how this new method could completely change the way AI is developed in many areas where there isn't enough data. Few-shot learning is the key to using AI in areas that weren't possible before, like personalized advice, healthcare diagnostics, and predicting rare events. As study in this area goes on, few-shot learning's potential to change the requirements for AI training and application development becomes clearer. This could lead to AI systems that are more efficient, flexible, and like humans. Keywords: Few-shot learning, Artificial intelligence, Meta-learning, Model generalizability, Catastrophic forgetting
INTRODUCTION: Traditional machine learning algorithms need very large datasets to learn well, which isn't always possible or useful, especially in specialized areas where data is hard to come by or expensive to buy. One study by Zhu et al., for example, found that deep learning models need a lot of labeled cases in each class to work well [1]. They did tests and found that a convolutional neural network (CNN) trained on a dataset of 1.2 million images did 95% of the ImageNet classification job correctly. However, the same model trained on only 10% of the data did only 85% correctly [1]. The F1 score of a named entity recognition (NER) model went from 0.75 to 0.87 when the training data was increased from 100,000 to 1 million sentences [2]. This is similar to Liu et al.'s natural language processing (NLP) work.
© 2024, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 209