A Note on Controlling the False Positive Rate in Image Classification

Abstract

While the usefulness of AI goes without saying, the development of reliable AI construction methods is essential to further accelerate the implementation of AI in society. For example, AI in the medical field is often used in primary screening to reliably exclude the majority of normal samples, and a high degree of trade-off is expected in making more error-free judgments with more samples. In this report, we introduce a new parameter to the loss function of the Siamese network of the deep distance learning model to enable mapping of hard-to-judgment samples to locations that reflect ambiguity among classes for the image classification problem. We also show that the mapping by introducing the new parameter provides a range in which samples that can be classified reliably can be extracted.