Online Learning of Image Recognition Task Offloading Policies Over Wireless Links

We study the problem of online learning of optimal offloading policies for image processing tasks, for minimizing a cost that is weighted sum of transmit energy and object recognition error rate. A mobile node generates image processing tasks that involve object recognition. There exist three options: (i) transmit the image to a remote server for processing with a deep-learning (DL) model, (ii) process locally with a simpler model, (iii) apply a lightweight, error-prone technique for object detection, and if objects are detected, then send image to the server. The proper offloading decision requires knowledge of the transmit energy cost and object recognition error rate for each option. However, these processes are non-stationary due to unpredictable object occurrence, mobility and propagation dynamics, and they depend on the object inference result which is unknown at decision time. We cast the problem as an adversarial multi-armed bandit, in which the EXP3 algorithm achieves sublinear regret. For the constrained problem, we propose an algorithm that extends EXP3 and achieves good regret in the objective and constraint, thus asymptotically learning the optimal static ran-domized offloading policy, while satisfying the error constraint. Performance is validated via numerical experiments informed by real-life object recognition measurements and models.