The document presents a bandit framework for improving video-based face recognition on mobile devices that face challenges from wireless transmission contention and cloud task congestion. It proposes a systematic learning method using multi-user multi-armed bandits to optimize transmission and resource allocation, demonstrating faster convergence than traditional reinforcement learning in dynamic environments. Performance is validated through simulations involving contention-based video streaming and face recognition algorithms under varying conditions.

1. BANDIT FRAMEWORK FOR SYSTEMATIC LEARNING IN WIRELESS VIDEO-
BASED FACE RECOGNITION
ABSTRACT
Video-based object or face recognition services onmobile devices have recently garnered
significant attention; giventhat video cameras are now ubiquitous in all mobile
communicationdevices. In one of the most typical scenarios for suchservices, each mobile device
captures and transmits video framesover wireless to a remote computing cluster (a.k.a. “cloud”
computinginfrastructure) that performs the heavy-duty video featureextraction and recognition
tasks for a large number of mobiledevices. A major challenge of such scenarios stems from the
highlyvarying contention levels in the wireless transmission, as well asthe variation in the task-
scheduling congestion in the cloud. Inorder for each device to adapt the transmission, feature
extractionand search parameters and maximize its object or face recognitionrate under such
contention and congestion variability, we proposea systematic learning framework based on
multi-user multi-armedbandits. The performance loss under two instantiations of theproposed
framework is characterized by the derivation of upperbounds for the achievable short-term and
long-term loss in theexpected recognition rate per face recognition attempt against the“oracle”
solution that assumes a-priori knowledge of the systemperformance under every possible setting.
Unlike well-known reinforcementlearning techniques that exhibit very slow convergencewhen
operating in highly-dynamic environments, the proposedbandit-based systematic learning
quickly approaches the optimaltransmission and cloud resource allocation policies based
onfeedback on the experienced dynamics (contention and congestionlevels). To validate our
approach, time-constrained simulationresults are presented via: (i) contention-based H.264/AVC

2. video streaming over IEEE 802.11 WLANs and (ii) principal-componentbased face recognition
algorithms running under varyingcongestion levels of a cloud-computing infrastructure.