Introduction and Background

In most science areas, which include DNA & molecule designs in micron level and earth environment sciences in macro level, it is so important to extract meaningful information from big data, which is superficially useless data with huge size. The extraction techniques are called data mining. Data mining is so costly that it is difficult to process it in traditional ways. To achieve much more efficient and accurate data mining and result in innovative science technologies, we have to propose new approaches based on mathematical theorems in algorithms and execution styles.

Division of Super Distributed Intelligent Systems, which is the previous division, especially focuses attention to medical and bio-systems, and has developed next generation data mining softwares together with researchers in artificial intelligence and statistics areas. In the process of that, we have found that we have to not only enhance parallelization/distribution and propose new approaches based on mathematical theorems to achieve new innovative technologies. In Division of Digital Transformation, we will improve the results of the previous division, and develop new big data processing manners based on performance and accuracy issues that the results have exposed. For example, we will continuingly enhance execution efficiency in the low level that is related with programming languages, parallel/distributed algorithms, and network protocols. In addition, we will design new deep learning manners based on adjusting their super-parameters based on combinatory theorem. Eventually, we will apply these techniques and models to several areas such as image processing, power systems, machine learning, robot systems, software engineering tools and so on, including data mining.

Research Hierarchy

Fig.1. Relations between research areas

As shown in Fig.1, we address the issues of big data processing in three hierarchical levels, “applications”, “fundamentals, and “theories” as follows:

1. Applications

In this level, members who are specialists of each applications investigate issues of the applications based on their expertise, propose approaches to solve the issues, and check validity of results given by the solution. In the process, they give new models based on characteristics of the applications, and develop systems implementing the models. The results given by the systems are validated in mathematical methods.

2. Fundamentals

In this level, members directly improve performance of fundamental techniques such as A.I. and machine learning, and  propose new approaches of them. The improvement of performance includes network performance in distributed systems and sensor networks, and learning performance of A.I. through parallel and distributed techniques. The new approaches includes improvements of parallelism in instruction level on GPU, improvement of accuracy of existing machine learning, and development of new machine learning model  based on biological systems. The fundamental techniques and systems developed in the level are validated in mathematical methods.

3. Theories

In this level, members give proofs of techniques with black box parts such as deep learning and machine learning. Furthermore, through knowledges obtained in the process, they propose new methods or system models.

Expected Effectiveness

Productions developed and knowledges found in each level can quickly be shared by all the levels. Because of that, we can give domain specific effective solutions. For example, we have developed a system for detecting distraction of drivers based on movement of eyes in the previous division. The system can expose cognitive distraction of drivers through A.I.’s integrating environment information and eye movement data. In the system, since A.I. has to process huge various sensor data, it requires parallel learning and inference algorithms, and their parallel or distributed execution. Thus, truly efficient and accurate big data processing is given by improvement of the system in multiple levels, which is achieved by cooperation between specialists in several areas.

We believe that the challenges of this division will give breakthroughs in many traditional techniques, and open new horizons for big data processing.

Fig.2. Mutual feedback between improvement and analysis

Future Development Goals

Development of big data processing manners that give accurate predictions for practical missions through mutually and spirally reflecting feedbacks between machine learning systems and statistical analyses of their results.

Message

This research division aims to give high-performance and more accurate big data processing manners with mutual feedbacks between machine learning systems and statistical analyses of their results, based on mathematical foundations in various levels.  The processing manners include redesign from theory and implementation of systems and analyses. We believe that the challenges of this division will open new horizons for big data processing.