Attention Consistency Regularization for Interpretable Early-Exit Neural Networks

Early-exit neural networks can now be both faster and more trustworthy: The paper demonstrates that by using attention consistency regularization, these networks maintain up to 98.97% accuracy while achieving nearly 2x faster inference speeds—meaning AI systems can make predictions quicker without sacrificing the quality of their decisions or explanations.

Consistency in AI explanations matters for trust: The research found that early-exit networks often focus on different features at different layers, which can produce conflicting explanations for the same prediction. By improving attention consistency by up to 18.5%, the framework ensures that users receive coherent explanations regardless of which layer makes the final decision, building confidence in the system's reasoning.

Interpretability and efficiency don't have to be trade-offs: Rather than choosing between understanding how an AI system works or making it run efficiently on resource-limited devices, this approach achieves both simultaneously—enabling explainable AI in practical, real-world applications like mobile devices or edge computing environments where computational power is limited.

The method uses attention maps as a transparency tool: By aligning visual attention patterns across different exit points in the network, the framework makes it clearer what features the AI is actually "looking at" when making decisions, providing a more transparent window into the model's decision-making process.