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.

What problem does this paper attempt to address?

The problem this paper attempts to address is that current deep learning technology in machine learning mainly remains at the surface learning level, lacking the conceptual understanding and knowledge transfer capabilities found in human deep learning. Specifically, the paper points out the following major limitations of current deep learning models: 1. **Interpretability**: Deep learning models are difficult to understand or interpret. 2. **Transferability**: The knowledge learned by deep learning models cannot be effectively transferred to new tasks. 3. **Context Adaptability**: Deep learning models cannot utilize contextual knowledge for adaptation. 4. **Large Amount of Labeled Data Requirement**: Deep learning models require a large amount of labeled data for training. To overcome these limitations, the authors propose Concept-Oriented Deep Learning (CODL), which aims to elevate machine deep learning to the level of human deep learning by introducing concept representation and conceptual understanding capabilities. The main aspects of CODL include concept maps, concept representation, concept instances, and a concept representation learning system, supporting incremental and continual learning.