Verification of the Effect of Label Modification on Neural Network Learning

This study focuses on the problem of subjective labeling that arises in the application of machine learning with neural networks to the medical field. The first problem is that of subjectivity, which arises because the labeling of medical data relies on expert knowledge and the process can be subjective. The second problem is the reduced information content of labels caused by discrete and quantitative labeling. If the intermediate values between the discrete values also have evaluative meaning, the detailed information is lost due to the discrete numerical labeling. We focus on such instability in labels and propose an approach to label modification using values inferred during neural network training. Specifically, we consider labels that are discrete, quantitative numbers as continuous numbers, compare the inferred value with the label value at each step in the learning phase, and compare the label value with the inferred value at each step in the learning phase. The update method is to give the label a sign-responsive variation. The goal is to modify each label using the value inferred by the inference made during learning in this way. As an experiment, we conducted an experiment on label modification under pseudo-conditions to verify its effectiveness. The results showed that when the amount of label modification was controlled by a constant, the generalization performance and the nature of label modification differed depending on the constant. In particular, generalization performance was improved when the constant of the label modification amount was determined so that the label value was separated from the inferred value at each step.