Abstract

Most state-of-the-art machine learning techniques revolve around theoptimisation of loss functions. Defining appropriate loss functions istherefore critical to successfully solving problems in this field. We present asurvey of the most commonly used loss functions for a wide range of differentapplications, divided into classification, regression, ranking, samplegeneration and energy based modelling. Overall, we introduce 33 different lossfunctions and we organise them into an intuitive taxonomy. Each loss functionis given a theoretical backing and we describe where it is best used. Thissurvey aims to provide a reference of the most essential loss functions forboth beginner and advanced machine learning practitioners.

 

Quick Read (beta)