Society for Science on Form, Japan, was founded in March 1985. The Society has been aiming to establish an interdisciplinary science centering on the concept of “form” (KATACHI) beyond the framework of conventional science classified by the subject of research.

In recent years, artificial intelligence (AI) has made remarkable progress. Clearly, the Society for Science on Form should also aim to develop the interdisciplinary science using AI as a tool. This special issue contains four interesting papers that serve as examples. They include deep neural network (DNN) to classify 3D structures, which can be used in engineering in general, and the use of biological signals, which is becoming indispensable in modern mobile terminals and is expected to develop in the future.

We hope that these four papers will contribute to the future developments of the Society for Science on Form, Japan.

High-density breasts are common among Japanese women, and thus, the detection rate of lesions is low, as they are obscured by the mammary glands during mammography screening. Therefore, attempts have been made to classify mammary gland concentration and present individuals with the risk of lesion obscuration; however, a large variability remains among doctors. In this study, we attempted to objectively classify mammary gland concentration by using deep learning. We examined the parameters of the learning image using a neural network console as a deep learning creation tool. As a result, a square image with a resolution of 64 × 64 and 100 learnings had the highest classification accuracy (88%). As the mammary gland concentration classification is a classification of the mammary gland of the entire breast, it is considered that the method by reduction has obtained a certain degree of accuracy. Next, training image data were expanded to obtain the subsequent classification accuracy. For data augmentation, the values of angle, brightness, and contrast were changed such that the image was similar to that before data augmentation. The resolution of 256 × 256 and network AlexNet were optimal, and 100% classification accuracy was obtained. In other words, the data augmentation—performed such that learning is equivalent in all classes—was considered effective. Conversely, because these are input images with similar features, they may not have various features. Therefore, verification using an unknown evaluation image will be required in the future.

Reverse engineering, defined in this study as obtaining a geometric three-dimensional (3D) computer-aided design (CAD) model from point cloud data, is often used in the manufacturing industry. However, it is necessary to predict the classes of primitive shapes for all the segments in a point cloud in order to generate the 3D CAD model surfaces needed, which is the primary factor driving the increase in required man-hours required for reverse engineering. Herein, we propose a simple deep neural network (DNN) that has relatively low temporal and spatial computational complexity in comparison to other models that can be used to classify point cloud segments. We then show how we trained our model on a dataset of partial point clouds containing primitive shapes to produce several point cloud size and dataset size combinations capable of achieving classification accuracy levels of over 90%. Our experimental results show that it is possible to classify point cloud segments with a small artificial dataset for use in reverse engineering processes.

It has been reported that the motion sickness is induced by visual approaches especially in the peripheral vision. Human visual information in peripheral is processed along dorsal streams in the brain. In this paper, we measured the radial motion of the young and the elderly while viewing video clips. We herein conducted the statistical machine learning (ML) to classify any test data into the peripheral and central vision. As a result, the values of the accuracy in the young were tended to be higher than those in the elderly. In addition, a comparison of the correct rate of the classification between peripheral and central vision showed that the rate for the central vision was higher than that for the peripheral vision.

Spectral analysis methods are commonly used for the Electrogastrography. However, spectral analysis is not sufficient to evaluate an electrogastrogram, and a certain size of time sequences is required for the analysis. Therefore, it is difficult to analyze time-series data over a relatively short period of time. The purpose of this study was to apply an analysis method using deep learning to the electrogastrograms of healthy young subjects in the seated posture during meal loading and in the supine posture after meal loading. The results of the analysis of the electrograms embedded in a low-dimensional feature space using an autoencoder showed that it was possible to estimate the state of the stomach before and after a meal in a short period of time.