Learning From Small Samples in the Age of Big Data

Abstract

Humans learn new concepts from a few observations with strong generalisation ability. Discovering patterns from small samples is complicated and challenging in machine learning (ML) and deep learning (DL). The ability to successfully learn and generalise from relatively short data is a glaring difference between human and artificial intelligence. Because of this difference, artificial intelligence models are impractical for applications where data is scarce and limited. Although small sample learning is challenging, it is crucial and advantageous, particularly for attaining rapid implementation and cheap deployment costs. In this context, this chapter examines recent advancements in small-sample learning. The study discusses data augmentation, transfer learning, generative and discriminative models, and meta-learning techniques for limited data problems. Specifically, a case study of convolutional neural network training on a small dataset for classification is provided. The chapter also highlights recent advances in many extensional small sample learning problems.