Concept-Oriented Deep Learning

TL;DR

Concept-Oriented Deep Learning (CODL) enhances traditional deep learning by integrating concept representations to improve interpretability, transferability, and contextual understanding, addressing key limitations of current models.

Contribution

This paper introduces CODL, a novel framework that incorporates concept graphs and representations to enable incremental and continual learning in deep neural networks.

Findings

Improves interpretability of deep models

Enhances transferability and contextual adaptation

Supports incremental and continual learning

Abstract

Concepts are the foundation of human deep learning, understanding, and knowledge integration and transfer. We propose concept-oriented deep learning (CODL) which extends (machine) deep learning with concept representations and conceptual understanding capability. CODL addresses some of the major limitations of deep learning: interpretability, transferability, contextual adaptation, and requirement for lots of labeled training data. We discuss the major aspects of CODL including concept graph, concept representations, concept exemplars, and concept representation learning systems supporting incremental and continual learning.