The Value of Symbolic Concepts for AI Explanations and Interactions

Modern Artificial Intelligence (AI), particularly deep neural networks, often lack transparency, making it difficult for human stakeholders to understand, evaluate, and correct model behavior—critical capabilities for mitigating shortcut behavior and ensuring trust in AI systems. This thesis addresses the need for trustworthy AI systems by investigating symbolic concept-based models that emphasize interpretability, revisability, and reliability in complex visual and logical reasoning. The thesis makes two central contributions: (i) identifying the potential of symbolic concept representations for building trustworthy AI, and (ii) developing frameworks that integrate neural and symbolic processing and leverage relevant inductive biases to learn these representations with minimal supervision. The first contribution establishes the value of symbolic concept representations through the Neuro-Symbolic eXplanatory Interactive Learning (NeSy-XIL) framework, demonstrating their role in enabling explainable and interactive AI. To address the challenge of learning these representations, the Motion and Object Continuity (MOC) framework is introduced, using temporal cues to derive meaningful object representations without strong supervision. Building on this, the thesis develops several frameworks—Neural Concept Binder (NCB), Interactive Concept Swapping Network (iCSN), Concept Bottleneck Memory Models (CB2M), and Pix2Code—each tackling different aspects of learning and revising symbolic object-based concept representations. Importantly, these frameworks exploit inductive biases, such as programmatic reasoning and object-factor disentanglement, to reduce reliance on supervision. Extensive evaluations demonstrate that these approaches achieve competitive performance compared to traditional neural models while enhancing interpretability, revisability, and reliability. These methods are particularly effective in mitigating shortcut behaviors and improving generalization. Additionally, the research highlights the benefits of combining both continuous and symbolic processing to create expressive yet human-accessible concept models. Finally, the thesis introduces synthetic datasets—V-LoL, CLEVR-Hans, ECR, and CLEVR-Sudoku—to benchmark and advance concept-based learning. These datasets offer diagnostic environments for assessing transparency, adaptability, and symbolic reasoning in complex relational and visual tasks. In summary, this thesis proposes that interpretable concept-based models grounded in neuro-symbolic learning can enable human users to understand and influence AI systems effectively. By bridging symbolic reasoning with flexible neural models, this work lays a foundation for trustworthy AI capable of learning, explaining, and adapting through symbolic concepts.