Abstract:Deep learning models have revolutionized numerous fields, yet their decision-making processes often remain opaque, earning them the characterization of "black-box" models due to their lack of transparency and comprehensibility. This opacity presents significant challenges to understanding the rationale behind their decisions, thereby impeding their interpretability, explainability, and reliability. This review examines 718 studies published between 2015 and 2024 in high-impact journals indexed in SCI, SCI-E, SSCI, and ESCI, providing a crucial reference for researchers investigating methodologies and techniques in related domains. In this exploration, we evaluate a wide array of interpretability and explainability (XAI) strategies, including visual and feature-based explanations, local approach-based techniques, and Bayesian methods. These strategies are assessed for their effectiveness and applicability using a comprehensive set of evaluation metrics. Moving beyond traditional analyses, we propose a novel taxonomy of XAI methods, addressing gaps in the literature and offering a structured classification that elucidates the roles and interactions of these methods. Moreover, we explore the intricate relationship between interpretability and explainability, examining potential conflicts and highlighting the necessity for interpretability in practical applications. Through detailed comparative analysis, we underscore the strengths and limitations of various XAI methods across different data types, ensuring a thorough understanding of their practical performance and real-world utility. The review also examines model robustness against adversarial attacks, emphasizing the critical importance of transparency, reliability, and ethical considerations in model development. A significant emphasis is placed on identifying and mitigating biases in deep learning systems, providing insights into future research directions that aim to enhance fairness and reduce bias. By thoroughly reviewing current challenges and emerging research directions, this article equips researchers with the knowledge and tools to advance the development of more transparent, fair, and reliable deep learning systems. Ultimately, this work aims to bridge existing literature gaps by offering a forward-looking perspective that fosters innovation and progress in the field. This comprehensive review not only illuminates the current state of XAI methodologies but also contributes to the broader understanding and enhancement of deep learning systems, ensuring their ethical and equitable application across various domains.

On reading and interpreting black box deep neural networks

Opening the Black Box of Neural Networks: Methods for Interpreting Neural Network Models in Clinical Applications

Analysis of Explainers of Black Box Deep Neural Networks for Computer Vision: A Survey

Illuminating the Black Box: Interpreting Deep Neural Network Models for Psychiatric Research

A Survey of the Interpretability Aspect of Deep Learning Models

Interpretable deep learning: interpretation, interpretability, trustworthiness, and beyond

A Comprehensive Review of Deep Neural Network Interpretation Using Topological Data Analysis

Analyzing Transformers in Embedding Space

Opening the Black Box: Analyzing Attention Weights and Hidden States in Pre-trained Language Models for Non-language Tasks

Toward Transparent AI: A Survey on Interpreting the Inner Structures of Deep Neural Networks

Why transformers are obviously good models of language

Learning Transformer Programs

Transparency and the Black Box Problem: Why We Do Not Trust AI

Interpretability of deep learning models: A survey of results

Representations as Language: An Information-Theoretic Framework for Interpretability

The Dark Side of the Language: Pre-trained Transformers in the DarkNet

Conversations that explain predictions

Unlocking the black box: an in-depth review on interpretability, explainability, and reliability in deep learning

Transformers are uninterpretable with myopic methods: a case study with bounded Dyck grammars

Interpreting Deep Learning Models in Natural Language Processing: A Review

Explainable Artificial Intelligence: Understanding, Visualizing and Interpreting Deep Learning Models