Bayesian Active Learning By Distribution Disagreement

The ever growing need for data for machine learning science and applications has fueled a long history of Active Learning (AL) research, as it is able to reduce the amount of annotations necessary to train strong models. However, most research was done for classification problems, as it is generally easier to derive uncertainty quantification (UC) from classification output without changing the model or training procedure. This feat is a lot less common for regression models, with few historic exceptions like Gaussian Processes. This leads to regression problems being under-researched in AL literature. In this paper, we are focusing specifically on the area of regression and recent models with uncertainty quantification (UC) in the architecture. Recently, two main approaches of UC for regression problems have been researched: Firstly, Gaussian neural networks (GNN) [6, 14], which use a neural network to parametrize µ and σ parameters and build a Gaussian predictive distribution and secondly, Normalizing Flows [16, 4], which are parametrizing a free-form predictive distribution with invertible transformations to be able to model more complex target distributions. Their predictive distributions allow these models to not only be trained via Negative Log Likelihood (NLL), but also to draw samples from the predictive distribution as well as to compute the log likelihood of any given point y. Recent works [2, 1] have investigated the potential of uncertainty quantification with normalizing flows by experimenting on synthetic experiments with a known ground-truth uncertainty. Intuitively, a predictive distribution should inertly allow for a good uncertainty quantification (e.g.