Adaptive Granular-Ball Concept-Cognitive Learning for Efficient and Robust Fuzzy Knowledge Representation in Classification Tasks

In many real-world scenarios, data unreliability and information-processing uncertainty seriously affect the credibility and reliability of artificial intelligence models. Probing a valuable cognitive learning paradigm is crucial for supporting and implementing the proposed artificial intelligence hypothesis. Concept-cognitive learning (CCL), as an emerging cognitive learning technology, investigates the knowledge-learning process using limited data through the lens of concepts. However, the existing CCL research has one prevalent disadvantage: instability and complex cognition. Thus, this work focuses on constructing an effective and interpretable cognitive learning paradigm for acquiring knowledge. Inspired by a multi-granularity knowledge representation, this paper proposes an adaptive and robust fuzzy knowledge representation paradigm, namely granular-ball concept-cognitive learning(GB-CCL). Unlike existing CCL systems, GB-CCL is a non-parameter cognitive learning system that achieves more accurate, faster adaptive capabilities for concept learning. The concept acquired through cognitive learning is structured conceptual knowledge, which is naturally interpretable and robust. Finally, extensive experiments verify that GB-CCL is an effective cognitive learning method with significantly better performance than state-of-the-art models in classification and robust learning.