Investigating Feasibility of Active Learning with Image Content on Mobile Devices Using ELM

This work investigates the feasibility of using the computational resources of a mobile device in active learning usage scenarios. It addresses the two main concerns, namely a way of fast model training or updating when more labels become available without re-training the whole Deep Learning model in case of image analysis, and the feasibility of running active learning workloads directly on a mobile device to improve the responsiveness and avoid using Cloud computing resources that become expensive at a large scale.

The results found that a mobile phone (Apple iPhone Xs in particular) is superior to CPU-bound workloads on a modern laptop. Two special discoveries relate to the latency of the first prection that turns out to be 20x faster on a phone, and some kind of short-lived acceleration after a user touches a phone’s screen that let small batches of up to 20 images to be processed twice faster than usual, in only 0.1 s for a classification of 20 images.