DART: Input-Difficulty-AwaRe Adaptive Threshold for Early-Exit DNNs

Researchers introduce DART, a new framework for early-exit deep neural networks that achieves up to 3.3x speedup and 5.1x lower energy consumption while maintaining accuracy. The system uses input difficulty estimation and adaptive thresholds to optimize AI inference for resource-constrained edge devices.

• →DART framework delivers up to 3.3x speedup and 5.1x energy reduction compared to static neural networks while preserving competitive accuracy.
• →The system introduces a lightweight difficulty estimation module that quantifies input complexity with minimal computational overhead.
• →Testing on AlexNet, ResNet-18, and VGG-16 demonstrates significant performance improvements across diverse DNN architectures.
• →Extension to Vision Transformers shows power gains but reveals accuracy trade-offs, highlighting need for transformer-specific optimizations.
• →The new Difficulty-Aware Efficiency Score (DAES) metric shows 14.8x improvement over baseline methods in balancing accuracy, efficiency, and robustness.