A Comparative Analysis on the Performance of Upper Confidence Bound Algorithms in Adaptive Deep Neural Networks

A new paper from researchers at Harokopio University of Athens presents a comparative analysis of four advanced Upper Confidence Bound (UCB) algorithms—UCB-V, UCB-Tuned, UCB-Bayes, and UCB-BwK—for Adaptive Deep Neural Networks (ADNNs) in edge computing environments. Edge devices face strict constraints on energy consumption and latency, making dynamic inference strategies critical. The team built on existing ADNNs that use the Multi-Armed Bandit (MAB) framework with the basic UCB1 strategy to dynamically select confidence thresholds for early exits, reducing computation without sacrificing accuracy.

The study evaluated these strategies on ResNet and MobileViT architectures using CIFAR-10, CIFAR-10.1, and CIFAR-100 datasets. Results show that all four new UCB strategies achieve sub-linear cumulative regret, meaning they efficiently balance exploration and exploitation. UCB-Bayes converged the fastest, followed by UCB-Tuned and UCB-V. Crucially, UCB-V and UCB-Tuned dominated the Pareto frontiers for accuracy-latency and accuracy-energy trade-offs, offering the best compromises for real-world deployments. This work provides a practical toolkit for optimizing AI inference on resource-constrained devices.