[Paper Review] Explainable AI: current status and future directions

Prashant Gohel, Priyanka Singh+1 more|arXiv (Cornell University)|Jul 12, 2021

Explainable Artificial Intelligence (XAI)55 references78 citations

TL;DR

A survey of Explainable AI (XAI) techniques across multimedia data (text, image, audio, video), comparing transparent and post-hoc methods, discussing benefits, limitations, and future directions for trustworthy AI applications.

ABSTRACT

Explainable Artificial Intelligence (XAI) is an emerging area of research in the field of Artificial Intelligence (AI). XAI can explain how AI obtained a particular solution (e.g., classification or object detection) and can also answer other "wh" questions. This explainability is not possible in traditional AI. Explainability is essential for critical applications, such as defense, health care, law and order, and autonomous driving vehicles, etc, where the know-how is required for trust and transparency. A number of XAI techniques so far have been purposed for such applications. This paper provides an overview of these techniques from a multimedia (i.e., text, image, audio, and video) point of view. The advantages and shortcomings of these techniques have been discussed, and pointers to some future directions have also been provided.

Motivation & Objective

• Motivate the need for explainability in AI for trust and transparency in critical applications.
• Summarize existing XAI techniques across multimedia data (text, image, audio, video).
• Discuss advantages and shortcomings of XAI methods and provide pointers to future research directions.
• Present a classification framework distinguishing transparent and post-hoc XAI approaches.
• Highlight applications in defense, healthcare, law, and industry to illustrate practical relevance.

Proposed method

• Classifies XAI techniques into transparent and post-hoc categories.
• Reviews model-agnostic and model-specific explanations and their applicability to different data types.
• Discusses strengths and weaknesses of saliency maps, LIME, SHAP, LRP, and provenance-based approaches.
• Outlines a classification tree for XAI methods with respect to data type (text, image, video, audio).
• Provides case studies and practical considerations for deploying XAI in real-world domains.

Experimental results

Research questions

• RQ1 What are the main XAI techniques proposed for multimedia data (text, image, audio, video) and how do they compare?
• RQ2 What are the advantages and limitations of transparent versus post-hoc XAI methods across different data modalities?
• RQ3 How can XAI contribute to trust, transparency, and fairness in critical applications such as healthcare, defense, and law enforcement?
• RQ4 What future directions and research opportunities exist to improve explainability in multi-modal AI systems?

Key findings

• XAI methods are categorized into transparent (intrinsically interpretable) and post-hoc (explaining after the fact) approaches.
• Saliency-based explanations, LIME, SHAP, LRP, and provenance/taxonomy methods provide varying degrees of interpretability and have trade-offs in fidelity and complexity.
• Model-agnostic methods (e.g., LIME, SHAP) offer broad applicability but face challenges such as instability and feature dependence.
• Counterfactual explanations (SEDC/SEDC-T) provide human-centric, contrastive justifications useful for regulatory contexts.
• Explainability concepts are extended across text, image, audio, and video, with specific challenges and opportunities per modality.