Improvement to Naive Loss Functions with Outlier Identifier

Abstract

We present a new Loss function, Loss with Outlier Identifier (LOI), a technique that produces a more robust calculation of prediction loss in Machine Learning fields and limits training time and procedures to minimum extend. LOI is designed based on the advantages of several well-known algorithm while compensating their disadvantages through interdisciplinary techniques. We show that by add two free parameters that do not require extra training, LOI is ensured to be continuous and derivable at all points and thus can be minimized through normal Gradient Descent algorithm. This function can be used to provide a more reliable loss for model training and thus produce a better model overall.