2021 年 25 巻 2 号 p. 53-61
Recently, deep neural networks (DNNs) have made outstanding progress and have been applied for solving various real-world problems. Among DNNs, convolutional neural networks (CNNs) are especially well known for their high performance as image classifiers. In this study, we generalize the CNN as a general classifier, propose a collaborative framework for neural networks and humans, and test the proposed collaborative model for medical diagnosis problems. Here, the CNN is not used as a stand-alone classifier but as a recommendation system, while humans make the final decision intuitively. To make this idea possible, a high-dimensional sample is first converted into a two-dimensional topological map before being subsequently given to the CNN as an input. The conversion of a general high-dimensional sample to a two-dimensional map allows the CNN to process it as an input while giving intuitive visual information for humans to be used in the final decision making. In this study, we attempt to expand the usage of CNNs for non-image inputs and propose a collaboration system for human and neural networks with the primary objectives of generating viable synergy between humans and AI. The paper is supported by some preliminary experiments to show the viability of our idea.
