The application of eXplainable artificial intelligence in studying cognition: A scoping review

Area of Science:

• Cognitive Neuroscience
• Artificial Intelligence
• Explainable AI

Background:

• Advancements in artificial intelligence (AI) necessitate trustworthy AI, leading to the rise of explainable AI (XAI).
• Recent neuroscience research highlights XAI's critical role in studying cognitive processes.
• Understanding cognitive function and dysfunction requires robust methods for interpreting AI models.

Purpose of the Study:

• To systematically review and analyze XAI methods applied to cognitive neuroscience.
• To identify prevalent XAI techniques used for investigating cognitive mechanisms.
• To develop a framework for applying XAI in cognitive neuroscience research.

Main Methods:

• Scoping review adhering to Joanna Briggs Institute and PRISMA-ScR guidelines.
• Searched major databases: MEDLINE, Embase, Web of Science, Cochrane, Google Scholar.
• Qualitative assessment involving two independent reviewers for data screening, extraction, and thematic analysis.

Main Results:

• Twelve experimental studies from the last decade were included.
• 75% of studies focused on normal cognition (perception, memory, etc.), 25% on impaired cognition.
• Intrinsic XAI (58.3%) was most common, followed by attribution-based (41.7%) and example-based (8.3%) methods; explainability was local (66.7%) or global (33.3%).

Conclusions:

• XAI methods offer predictive power and robustness in cognitive neuroscience.
• Limitations include oversimplification, confounding factors, and inconsistencies.
• Future research needs to address causality and reproducibility challenges in XAI applications.