Loss Function Selection in NN based Classifiers: Try-outs with a Novel Method

Abstract

Neural Network (NN) based classifiers are frequently used in different disciplines like pattern recognition, image classification, regression, etc. The optimized NNs are produced to achieve more robust classifiers in terms of preventing the fluctuations at error by yielding better convergence. However, in the literature, an extensive study is not available that examines one of the most important dynamics of NNs: loss function (error metric) of algorithm. In this study, ten different loss functions are evaluated to find out which one stays more coherent to utilize in NN based classifiers. For this purpose, a novel optimized classifier is handled using Gauss map based chaotic particle swarm optimization (GM-CPSO) with NN. GM-CPSO-NN classifiers including ten different loss functions are compared on two challenging tasks (Parkinson's disease recognition and epileptic seizure detection). Four metric based comparisons and 2-fold cross validation method are processed to objectively test the performance of classifiers. As a result, it's concluded that GM-CPSO-NN comprising mean square error (MSE) attains remarkable performance for both tasks, and MSE is revealed as the most appropriate error metric to be considered for the datasets including low or high pattern numbers.