Dynamic Data Distribution-based Curriculum Learning

Curriculum learning has proven effective in enhancing the performance of a classifier by gradually training models on samples that range from simple to difficult based on prior information. We have previously explored the innovative curriculum learning approach known as Data Distribution-based Curriculum Learning (DDCL). In this study, we propose a novel extension to DDCL termed Dynamic DDCL, leveraging self-paced learning to create a more informed learner. Its dynamic curriculum promotes adaptive learning capabilities by adapting to the needs of the model as it evolves during training. We further introduce DDCL Ensemble, an ensemble learner that aggregates the enhancements of the distinct scoring methods present in DDCL and Dynamic DDCL. We assess the effectiveness of Dynamic DDCL using classifiers based on neural networks. The performance of DDCL Ensemble is evaluated against a counterpart ensemble learner which is devoid of any curriculum learning. Experimental findings highlight the superior performance and generalisation capabilities achieved by Dynamic DDCL and DDCL Ensemble, with performance increases ranging from 1% to 34% and 1% to 11% respectively, when compared to other self-paced learning methodologies and standard ensembles. In addition, they show potential in advancing the state-of-the-art in classifier optimisation for domains where training data is limited.