Boosting AI Trust: Interpretable Early-Exit Networks with Attention Consistency
Analysis
Key Takeaways
- •The paper introduces Explanation-Guided Training (EGT), a multi-objective framework for interpretable early-exit neural networks.
- •EGT uses attention consistency loss to align attention maps across different exit points, improving interpretability.
- •Experiments show EGT achieves comparable accuracy with faster inference and enhanced attention consistency compared to baseline models.
“Experiments on a real-world image classification dataset demonstrate that EGT achieves up to 98.97% overall accuracy (matching baseline performance) with a 1.97x inference speedup through early exits, while improving attention consistency by up to 18.5% compared to baseline models.”