Gaussian Switch Sampling: A Second Order Approach to Active Learning

In active learning, acquisition functions define informativeness directly on the representation position within model manifold. However, for most machine learning models (in particular neural networks) this is not fixed due to training pool fluctuations in between rounds. Therefore, several popular strategies are sensitive experiment parameters (e.g. architecture) and do consider robustness out-of-distribution settings. To alleviate issue, we propose a grounded second-order definition of information content sample importance context learning. Specifically, by how often network “forgets” during - artifacts second order shifts. We show that our produces highly accurate scores even when representations constrained lack data. Motivated analysis, develop Gaussian Switch Sampling ( GauSS ). setup agnostic robust anomalous distributions with exhaustive experiments three in-distribution benchmarks, different architectures. report an improvement up 5% compared against four query strategies. Our code available at https://github.com/olivesgatech/gauss .