Explainable AI for Embedded Systems Design: A Case Study of Static Redundant NVM Memory Write Prediction

XAI can validate ML model decisions in embedded systems: The study demonstrates that explainable AI methods successfully identified known causes of silent stores (redundant memory writes), such as operations writing zero constants, confirming that XAI explanations align with established technical knowledge in the field.

Silent store elimination offers real performance and energy benefits: By identifying and removing redundant memory writes through ML prediction, the research shows potential improvements in both performance and energy efficiency—particularly important for systems using emerging non-volatile memory technologies where write operations are costly.

Model-agnostic XAI methods are practical for hardware design: The paper successfully applied two state-of-the-art model-agnostic XAI techniques to explain ML predictions in embedded systems, suggesting these general-purpose explanation methods can be effectively adapted to specialized technical domains beyond their original applications.

XAI helps catch ML model failures: The methodology revealed cases where silent stores were absent in loop induction variable operations, demonstrating how XAI can identify when ML models correctly don't make predictions—helping engineers spot potential blind spots in model behavior.

Transparency in ML-assisted hardware design is achievable: This case study provides a practical framework showing that machine learning models used in embedded systems design can be made interpretable and trustworthy, laying groundwork for broader adoption of responsible AI practices in hardware optimization.