A Survey on Incremental Learning

Abstract

Abstract:Incremental learning is one of the research hotspots in machine learning. In this paper, we view the complex changes of data as three changes that are the change of sample, the change of class and the change of feature, and analyze the popular machine learning classification algorithms which support incremental learning. And then we focus on reviewing the research of three types of incremental learning: Sample Incremental Learning, Class Incremental Learning and Feature Incremental Learning. Finally, we make a prospect on the focus and difficulty of future research of incremental learning.

Convolutional auto-encoded extreme learning machine for incremental learning of heterogeneous images

“…Table 2 (case 1) and Figures 3(a With adding new classes: We also compared CAE-ELM against Learn++ and ELM++ methods with MNIST dataset under scenario where new classes gets added in every new arriving batch of data. We split the MNIST dataset into three sets S1, S2, S3, where S1 trains classes (0-2), S2 trains classes (3)(4)(5) and S3 holds samples from classes (6)(7)(8)(9). Table 2 (case 2) and Figure 3 tabulates the results obtained.…”

“…Intelligent adaptation methods deal with the model complexities by reallocating the resources whenever the limit is reached. e) Efficient learning models: the incremental learning models [8] must be efficient enough to deal with the constantly arriving data. Even in limited resources, these models must deal with the newly added data by storing the information provided by the observed data in compact form.…”