With the advancement of deep neural networks (DNNs), their surprising performance has sparked widespread interest in exploring the feasibility of their applications across various fields. However, the inherent opacity of DNNs presents a significant barrier to their application in high-stakes fields where demonstrating the inference process of DNNs is as critical as the accuracy of model outputs. Although various methods, including explanation generation, transparent models, and human-model interaction methods, have been proposed to enhance explainability of DNNs for revealing the inference process, their effectiveness in revealing causality, correcting model bias, and guiding system design with multiple interaction scenarios remains limited. To address this research gap, this thesis proposes a concept-based causal explanation generation workflow, as a post-hoc explanation generation method, to illustrate the causality between interested concepts and model outputs with multiple formats of explanations through a global variational autoencoder probe, a causal structure discovery algorithm, a causal effect estimation method, and a set of explanation generation tools. To correct model bias, a causal-guided model fine-tuning workflow is proposed, consisting of causal-based sample selection, concept-based low-rank fine-tuning module, and a do-calculus optimisation strategy for implementing causal inference in DNNs. This workflow converts black-box DNNs into transparent models and corrects model bias through model fine-tuning. Last, a human-in-the-loop framework with explanations is proposed to guide the DNN-based system design and validation on multiple human-model interaction scenarios. In experiments, the proposed workflows outperform state-of-the-art methods in enhancing DNN explainability and correcting model bias in computer vision and natural language processing tasks. Additionally, a case study of pest management in agriculture is completed, demonstrating the feasibility of the proposed frameworks for establishing explainable DNN-based systems.
