Stratified Active Learning

Active learning is a method of training a machine learning model using less data by carefully selecting the most informative samples (read more here).

Generally, active learning is evaluated in terms of global training efficiency i.e. the discriminative model performance given the number of samples. This is an important aspect of active learning but, I think this often fails to properly evaluate the generalisability of performance across-domains (e.g. locations, time, classes etc).

This pre-print explores the generalisability of model performance across spatial and temporal scales and investigates how active sampling and diversification methods can improve the generalisability of model performance across domains.