Concept-Level Cognitive Modeling via Multimodal Interaction: A Neuro-Symbolic Hybrid Approach for Early Concept Acquisition

Concept-level cognitive modeling is fundamental to building intelligent systems that can interpret, reason, and interact with the world in a human-like manner. While recent advances in deep learning and multimodal representation have enabled impressive progress in pattern recognition, they fall short of capturing abstract, structured concepts such as spatial relations, agency, and containment that underpin human cognition. This paper presents a neuro-symbolic hybrid framework for early concept acquisition through multimodal interaction. Our approach integrates perceptual input from vision and language with symbolic abstraction, enabling the system to form and refine concept representations via active cross-modal grounding. Central to our method is a dynamic concept representation layer that links symbolic units to neural embeddings, allowing the model to incrementally build a structured concept space without requiring predefined ontologies or exhaustive retraining. An interactive learning loop drives the alignment between modalities, supports hypothesis testing, and facilitates continual concept refinement. We validate our framework on both synthetic and real-world datasets across a range of conceptual tasks including spatial relation understanding, object function reasoning, and affordance inference. Experimental results demonstrate that our model outperforms purely neural and purely symbolic baselines in generalization, interpretability, and adaptability. This work contributes to the development of cognitively inspired AI systems that learn abstract knowledge from interaction, bridging the gap between subsymbolic perception and symbolic reasoning.