Explainability vs. Performance: Bridging the Trade-Off in Deep Learning Models

Abstract

This work explores the explainability-performance trade-off in deep learning models, especially in life-and-death scenarios such as autonomous driving and medical diagnostics. The more complex and integrated AI models become and the more integrated they are into safety-critical systems, the greater the need to ensure their transparency to preserve trust and accountability. It discusses the different ways to balance model performance and explainability, both in the form of interpretability methods and post-hoc techniques of explanation such as LIME and SHAP. Tesla with Autopilot and IBM Watson Health are just one example of case studies where this trade-off is challenged and its impact is demonstrated. It reveals that to make the use of AI systems safe and ethical, high performance is necessary, and the decisions should be made transparently. The study adds to the body of existing knowledge by highlighting the need to be more transparent with AI and provides a recommendation to enhance model interpretability without making the model too inaccurate. This paper highlights the current necessity to see further developments in explainability of AI, especially in areas that are vital to safety.