Recently, reinforcement learning models have achieved great success, mastering complex tasks such as Go and other games with higher scores than human players. Many of these models store considerable data on the tasks and achieve high performance by extracting visual and time-series features using convolutional neural networks (CNNs) and recurrent neural networks, respectively. However, these networks have very high computational costs because they need to be trained by repeatedly using the stored data. In this study, we propose a novel practical approach called reinforcement learning with convolutional reservoir computing (RCRC) model. The RCRC model uses a fixed random-weight CNN and a reservoir computing model to extract visual and time-series features. Using these extracted features, it decides actions with an evolution strategy method. Thereby, the RCRC model has several desirable features: (1) there is no need to train the feature extractor, (2) there is no need to store training data, (3) it can take a wide range of actions, and (4) there is only a single task-dependent weight parameter to be trained. Furthermore, we show the RCRC model can solve multiple reinforcement learning tasks with a completely identical feature extractor.

We investigate prediction accuracy for time series of Echo state networks with respect to several kinds of activation functions. As a result, we found that some kinds of activation functions with an appropriate nonlinearity show high performance compared to the conventional sigmoid function.

No abstract

Machine learning, applied to medical data, can uncover new knowledge and support medical practices. However, analyzing medical data by machine learning methods presents a trade-off between accuracy and privacy. To overcome the trade-off, we apply the data collaboration analysis method to medical data. This method using artificial dummy data enables analysis to compare distributed information without using the original data. The purpose of our experiment is to identify patients diagnosed with diabetes mellitus (DM), using 29,802 instances of real data obtained from the University of Tsukuba Hospital between 01/03/2013 and 30/09/2018. The whole data is divided into a number of datasets to simulate different hospitals. We propose the following improvements for the data collaboration analysis. (1) Making the dummy data which has a reality and (2) using non-linear reconverting functions into the comparable space. Both can be realized using the generative adversarial network (GAN) and Node2Vec, respectively. The improvement effects of dummy data with GAN scores more than 10% over the effects of dummy data with random numbers. Furthermore, the improvement effect of the re-conversion by Node2Vec with GAN anchor data scores about 20% higher than the linear method with random dummy data. Our results reveal that the data collaboration method with appropriate modifications, depending on data type, improves analysis performance.

Reservoir computing derived from recurrent neural networks is more applicable to real world systems than deep learning because of its low computational cost and potential for physical implementation. Specifically, physical reservoir computing, which replaces the dynamics of reservoir units with physical phenomena, has recently received considerable attention. In this study, we propose a method of exploiting the dynamics of road traffic as a reservoir, and numerically confirm its feasibility by applying several prediction tasks based on a simple mathematical model of the traffic flow.

No abstract

Collecting large amounts of data is beneficial in machine learning to generate models that are less biased. There are many cases in which pieces of similar data are distributed among organizations, and it is difficult to integrate these data owing to issues involving privacy and cost. Integrating these distributed data without delivering the original data leads to the concept of data collaboration, which combines data held by different organizations in a secure manner. We propose a method in which a distance matrix of the original data obtained using common data among organizations is shared to learn neighbor information of the original data. Specifically, the proposed method robustly integrates distributed data, which is of as good quality as connected raw data, in cases where the amount of data in each organization is small and the data bias is large. In addition, the proposed method is applicable to data contaminated by noise. To demonstrate the effectiveness of the proposed method, we performed a classification task on open biological data divided into several pieces and found that the classification results for divided data were as precise as when all data were available. Finally, we show that the robustness of the method against noise improves the anonymity of the original data as a by-product.

No abstract