This paper proposes a system that encourages users to discover disciplines such as mathematics and biology, which they have not explored before. The recommendation of unexpected information is highly demanded in current recommender systems, which are used in various situations. To meet this need, the present study considers the serendipitous discovery of unexpected disciplines related to the information of interest to users. Serendipity means “the ability to see what others do not, and to discover useful things with your own wisdom.” Our system encourages users first to select a photograph of interest, and then to ponder the relevance of another photograph appearing on the display. The system is also equipped with a hint generation function that outputs an academic discipline related to the photos in order to foster users’ dictionary of unexpected relationship between the two photos. Whether users gain serendipity from the system was evaluated in an experiment. The results confirmed that the system induced serendipity by providing unexpected information through photographs linked to disciplines.

Using the Apriori-based method proposed by Agrawal for transactional data processing, we are proceeding with rule generation and construction of its execution environment from tabular data sets. This time, we will reconsider the functionality of decision-making by using the rules obtained. The application of rules to decision support is a long-standing problem, and we think that the need for it is increasing as a method to complement the recent black-boxing of conclusions in AI. Some IT companies have recently released software tools on “explainable AI” for black-boxing. In this paper, based on the presentation at FSS2020, we consider the functionality that can fully explain the conclusion using rules.

In order to evaluate quality of a steel weld joint, radiographic examination is utilized in addition to ultrasonic examination. The radiographic examination is a method of projecting cavities and foreign materials in the test object onto the film by utilizing the property that the radiation penetrates the test object and exposing film. Since the quality evaluation of the obtained film image is performed by human eyes, it depends largely on the empirical rule, and there is a problem that it takes a lot of time to learn and oversight occurs. In this research, we consider a method to automatically detect flaws from the radiographic image using AI (CNN).

The Convolutional Neural Network (CNN) is applied to several applications and is expected to the embedded systems such as IoT devices. However, these systems need high calculation costs and high power consumption. Because in general, those system requires the using GPUs and its hard to implement on the small embedded system. In recent years, FPGAs have been applied to the auto-control systems, defect inspection systems, and the security systems. Especially applying for the image processing technologies for industrial products were adapted. Hardware acceleration is one of the techniques to improve processing speed and is often used for image processing. Our work has designed the hardware acceleration for CNN and compared it to software processing. The hardware-based modules of CNN were implemented on FPGA and tested. This paper shows the details architecture, was designed in our research, and the verification results of the real-time processing.

In this study, we consider generating efficiently labeled data from map symbols of GIS data as a means to efficiently increase the data. Further, we propose to perform semi-supervised learning using this method. We demonstrate the effectiveness of the proposed method in 6 classes of land-use classification.

Recently, interpretability of classifiers has been actively discussed in some real-world applications like medical and financial fields. Michigan-type Fuzzy Genetics-based Machine Learning (FGBML) is one of the most-well known methods for generating fuzzy classifiers with high interpretability. However, many real-world classification problems require high classification performance of a specific class which is less frequent than the others. For those problems, fuzzy classifiers generated by FGBML often have low classification performance for minority classes. Therefore, in this study, we extend FGBML for minority classes by applying the following four operations to the conventional Michigan-type FGBML: i) weighting the rule evaluation function by the number of patterns for each class, ii) changing the rule set evaluation function to the harmonic mean of the classification accuracy of each class, iii) selecting the base pattern in the heuristic rule generation based on the classification accuracy of each class, and iv) weighting the consequent part of each rule by the ratio of the number of the patterns between classes. Through computational experiment, we examine the effect of each change on the classification performance and show that the proposed method generates a fuzzy classifier with better classification performance than the conventional FGBML with SMOTE.

Multi-objective fuzzy genetics-based machine learning for multi-label classification called MoFGBML_ML is a classifier design method for interpretable fuzzy classifiers. It generates a number of non-dominated fuzzy rule-based classifiers with different accuracy-complexity tradeoffs. In multi-label classification, some performance metrics have been simultaneously used for comparison. However, MoFGBML_ML can handle only one performance metric in a single run. In this paper, we extend two-objective MoFGBML_ML to many-objective optimization. In the many-objective optimization formulation, we use several performance metrics as objective functions simultaneously. This extension enables MoFGBML_ML to obtain multiple optimal classifiers with respect to several performance metrics for multi-label classification in a single run.

In a multi-modal multi-objective optimization problem, there exist some Pareto optimal solutions for any point on the Pareto front. A multi-modal multi-objective evolutionary algorithm needs the abilities of approximating better both the Pareto front and the Pareto set. However, most of existing multi-modal multi-objective evolutionary algorithms use the population convergence in the objective space as the primarily evaluation criterion. As a result, they do not always have a high approximation ability of the Pareto set in the decision space. To approximate better both the Pareto front and the Pareto set, we propose a decomposition-based multi-modal multi-objective evolutionary algorithm. In our proposed algorithm, a multi-modal multi-objective optimization problem is transformed into a number of two-objective subproblems. In each subproblem, solutions are optimized in terms of the corresponding scalarizing function and the decision space diversity. The proposed algorithm can maintain not only solutions with good convergence to the Pareto front but also diverse solutions in the decision space. Experimental results show that the proposed algorithm has a high approximation ability to both the Pareto front and the Pareto set.

Clustering algorithms can flexibly extract useful knowledge from input data. Therefore, clustering algorithms are often applied to data preprocessing such as dimensionality reduction and feature extraction. Clustering algorithms can also be applied to classifiers thanks to a good knowledge extraction ability. As a conventional study of applying clustering algorithms to classifier design, the algorithm has been proposed that explicitly learns decision boundaries by applying a clustering algorithm to each class of data. However, there are some problems such as instability of learning and slow processing. In this study, we propose a classifier by utilizing the Fast Topological CIM-based Adaptive Resonance Theory (FTCA) that achieves both excellent self-organization performance and high-speed learning. Our experimental results in this paper show that the proposed algorithm has better classification performance compared to other clustering-based classifiers.

We examine a possibility of personal authentication using forearm surface EMG (s-EMG) during gesture operation. The s-EMG was measured 10 times from the 5 subjects during performing 6 gestures. In our previous research, we performed the gesture identification by using Support Vector Machine (SVM). As the feature values, we used the maximum and minimum values and their times of the time domain data during gesture. As the result, the identification rate was 66.7%. In this paper, to improve the identification rate, we increased the future values which we used, and we introduced the selection algorithms of the important feature values based on the random forest. As the result, the identification rate by using SVM was improved to over 80%. Moreover, the previous research claimed that the feature values in frequency domain was not effective. We found that some feature values in the frequency domain was effective for the gesture identification.

This paper deals with a quantitative evaluation of the color distribution at station streets and these effects on landscape impressions. The authors took videos during walking along the station streets, and these color distributions were quantified using the L^*a^*b^* color system. We also conducted an experiment in which participants watched these videos and then reported their subjective landscape impressions using the SD method. As a result, we identified that there are two types of impression factors, stimulus and spatiality, which are influenced by the chroma of the landscape and spatial distribution of gaze, respectively.

The demand for applying AI has increased to automate immune cell activity analysis in recent years. One of the reasons is the shortage of analysts. Also, the work requires a lot of time and effort because the immune cell’s activity is confirmed and analyzed while advancing the video frame by frame. Therefore, if cell images can automatically select using an AI classifier, the work time becomes short. In this research, we aim to shorten these tasks. For the AI classifier, this research used CNN, which is a kind of deep learning methods and conducted a study using cell images in advance. From the results of the study, it was confirmed by experiments that CNN exhibited good recognition performance. Next, we defined a “Recognition Frequency Space” for calculating the cumulative frequency of highly recognized regions and automatically selecting immune cells. By this method, we were able to generate detailed location information that is recognized as a cell. Using these, we were able to accurately cut out 10 immune cells from the frame image of the actual moving image. In this report, we show that multiple immune cells could be cut out automatically.

We improve the performance of a road marking detection system by incorporating You Only Look Once (YOLO) into the processing for vehicle location estimation in autonomous driving technology. The conventional detection method uses a template matching process based on luminance values to detect road marking. However, there are some markings that cannot be detected by this method due to halation by sunlight or strong blurred of road markings. In contrast, the proposed method uses YOLO to search for areas where road marking exists and restricts the area of adaptation for template matching. Owing to this area restriction, the proposed method can prevent the occurrence of false detection, lower the detection threshold for template matching, and reduce the number of previously undetected road markings. In addition, the search area for template matching is restricted, which also can improve the processing speed. Experimental results show that the proposed method is able to reduce the number of undetected road markings compared to the conventional method while keeping the number of false detections to zero. The accuracy of the system was improved by 0.013 and the processing speed was increased by 4.6 FPS compared to the previous method.

In recent years, research and development on smart houses that intelligently control various indoor devices (lighting, air-conditioning, etc.) have been active due to advances in communication and machine learning technology. In this paper, we propose and implement a system to realize novel smart house. The proposed system has two major features. One is various devices required for the system (sensing, processing, and actuation). It is an architecture called “distributed smart house” that disperses and arranges the distributed inference devices. In the other, the reasoners necessary for intelligent control are distributed within the device, local network, and cloud network. It is a “multilayer AI” that is implemented by dividing it into multiple layers. Due to these features, it is possible to construct a smart home system flexibly and efficiently. As for the system to be implemented, we used voice recognition using Google Cloud Speech and human posture recognition using Jetson TX2 placed in the local network. In addition, we used ROS as the software platform. This makes it possible to easily expand future functions and improve performance.

In this paper, we examine the amount of speech of a robot that becomes a walking partner with a person on a purposeless walk. The control of the volume of speech is a factor that contributes greatly to the impression of the listener on the speaker and on the conversation itself. In walks, it is very important to obtain knowledge about the amount of utterance, because the conversation with the partner who walks with them contributes to the impression of the walk itself. In this paper, we conducted an intra-participant verification experiment with four kinds of speech volume, and investigated the sensation of more or less speech volume and the change of impressions with speech volume. The results showed that the amount of utterance was felt to be more when the time between the start of the utterance and the start of the next utterance was 10 seconds and less when the time was 40 seconds in a short sentence of 4–5 seconds. When the time between the start of one utterance and the start of the next was less than 20 seconds, they rated significantly higher for the other conditions in terms of their reputation as a walking partner, their likability, and their impression of whether they were intelligent or not.