Ontology-Powered Hybrid Extensional-Intensional Learning
Alvis C. M. Fong, G. Y. Hong · 2019
Deep learning has made headlines in the past few years due to successes in tasks, such as self-driving vehicles and board games, which were previously thought difficult or impossible. The successes have generated much interest in artificial intelligence among researchers and members of the public. However, deep learning algorithms generally require very large labelled data sets to work well and large labelled data sets are not always readily available. In addition, most machine learning techniques, including deep learning, often perform well statistically but can fail miserably when, for example, data are deliberately perturbed in an adversarial attack. Another criticism of deep learning techniques is a relative lack of explainability. This paper proposes the use of intentional learning to simultaneously address these issues. Preliminary evaluation on the MNIST data set has shown promising results. Specifically, by combing extensional and intensional learning, it is possible to achieve similar accuracy result as extensional learning only using only one-sixth of the original training data set.