Losses for Deep Probabilistic Regression

Franco Marchesoni-Acland, Andrés Herrera, Rodrigo Alonso-Suárez, Jean-michel Morel, Josselin Kherroubi, Gabriele Facciolo

TL;DR: A review of losses for deep probabilistic regression

Abstract: Probabilistic regression is used in fields such as healthcare, finance, energy, robotics and meteorology. Although many works have dealt with probabilistic regression, they have frequently done so independently, often failing to compare against each other. This paper reviews probabilistic regression and aims at providing a unified overview of the area. We experimentally compare diverse approaches and observe that direct methods perform comparably to their sample-predicting counterparts, while being simpler to train and cheaper to infer with. We then introduce a taxonomy that sheds light onto the design choices behind each of the direct methods, suggesting new ones. The main takeaway is that simple methods can serve as strong baselines and should not be disregarded.

Primary Area: Probabilistic Methods->Everything Else

Keywords: probabilistic methods, probabilistic regression, deep learning, review, survey

Submission Number: 8285