CD-SLFN: A Curiosity-Driven Online Sequential Learning Framework with Self-Adaptive Hot Cognition
Chenyu Sun, Chunyan Miao · 2022 International Joint Conference on Neural Networks (IJCNN) · 2022
In human-level artificial intelligence, curiosity can guide agents to actively explore and learn with limited prior knowledge. The existing literature has attempted to leverage curiosity in online sequential learning problems; however, they are computationally expensive in general and lack generalization capability. In this paper, we present a unified curiosity-driven learning framework to tackle online classification problems. Based on the psychological theory of human curiosity, five collative variables including novelty, uncertainty, complexity, surprisingness, and change, are quantitatively characterized to model artificial curiosity. We seamlessly integrate these collative variables into a regularized single-hidden-layer feedforward neural network (SLFN) to encourage curiosity-driven learning. As a result, the proposed curiosity-driven SLFN (CD-SLFN) can actively select the most representative data in an online sequential manner, and flexibly adapt the model complexity to prevent overfitting. Compared to other online classifiers, CD-SLFN empowered by intrinsic motivation has demonstrated superior generalization capability, especially in the early learning phase with only limited data. The experimental results on various datasets indicate that the proposed CD-SLFN can consistently outperform other SLFN baselines, with the desired learning capability through self-adaptive hot cognition.