Comparative Study of UCB Algorithms in Adaptive Deep Neural Networks

A new arXiv preprint introduces four additional Upper Confidence Bound strategies for Adaptive Deep Neural Networks (ADNNs) in edge computing. The study compares UCB-V, UCB-Tuned, UCB-Bayes, and UCB-BwK against the standard UCB1, focusing on trade-offs between accuracy, energy consumption, and latency. Edge environments impose strict constraints, making adaptive inference critical. The work builds on the Multi-Armed Bandit framework to dynamically select optimal confidence thresholds for early exits. This is the first comparative analysis of these UCB variants in ADNNs.