Machine learning with insufficient data for classification of food oil mixtures by means of LIF spectra

Abstract In recent years, the algorithms of artificial intelligence (AI) attract increasing attention of scientists since they open doors to efficient solutions of many problems that otherwise require a lot of time, effort, expenses and often inspiration. A main challenge to their wider application in biophotonics is the lack of ample amount of diverse and representative data for training. Therefore, we present here the application of Neural Network (NN) algorithms trained with insufficient data for solving a task related to the classification of mixtures of food oils (sunflower and extra-virgin olive oil) with different concentrations for verification of food quality. The task is approached both as a classification and as a fitting task. The cases of training with raw Laser-Induced Fluorescence (LIF) data and with diagnostic parameters with biochemical meaning, and also when extending the insufficient datasets with and without adding noise are reported and compared.