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

A new study compares five Upper Confidence Bound strategies for Adaptive Deep Neural Networks, identifying which algorithms best balance accuracy, latency, and energy use in edge computing environments. The research, led by Grigorios Papanikolaou, builds on Adaptive Deep Neural Networks (ADNNs) that use the Multi-Armed Bandit framework to enable early exits during inference [1]. While prior work relied solely on the UCB1 strategy, this study introduces four additional variants: UCB-V, UCB-Tuned, UCB-Bayes, and UCB-BwK [2]. The algorithms were tested on ResNet and MobileViT architectures using the CIFAR-10, CIFAR-10.1, and CIFAR-100 benchmark datasets [2]. Edge computing environments impose strict constraints on energy consumption and latency, making the deployment of deep neural networks a significant challenge [2]. Smart, adaptive inference strategies that dynamically balance computational cost with predictive accuracy are therefore critical in these scenarios [2]. The Multi-Armed Bandit framework treats each possible confidence threshold as an arm, with the UCB algorithm selecting thresholds that optimize the trade-off between early exiting and accuracy over time. Experimental results showed that all five strategies achieved sub-linear cumulative regret, meaning their performance loss relative to an optimal oracle grew slower than linearly over time [2]. UCB-Bayes converged the fastest, followed by UCB-Tuned and UCB-V [2]. When examining Pareto Frontiers—the set of solutions where no single metric can be improved without degrading another—UCB-V and UCB-Tuned dominated the accuracy-latency and accuracy-energy trade-offs [2]. The study's approach to adaptive inference parallels broader trends in machine learning where ensemble methods and dimensionality reduction techniques have long been used to manage computational complexity. Random forests, for instance, correct for decision trees' tendency to overfit by averaging predictions across multiple trees, a concept introduced by Tin Kam Ho in 1995 and later extended by Leo Breiman and Adele Cutler [3]. Similarly, principal component analysis transforms data onto new coordinate systems to identify directions capturing the largest variation, enabling dimensionality reduction in fields from population genetics to atmospheric science [4]. The implementation code for the UCB comparison is publicly available on GitHub [2]. The paper was submitted to arXiv on April 27, 2026, and last revised on May 22, 2026 [1].