Jump to contents

 T-SOUL

Vol.20 Artificial Intelligence Discovering Values of IoT Data -Analytics towards Deep Learning-

Print

#02 Providing a Framework for Automatic Learning Platform Processes from Data Augmentation to Learning are Automated Challenge: Higher Accuracy Achieved with Shorter Time and Reduced Cost Shinichi Kashimoto Chief Specialist, Deep Learning Technology Department IoT Technology Center Industrial ICT Solutions Company, Toshiba Corporation

Toshiba is providing an optimal and high-accurate deep learning environment as a set tailored to the respective purposes of customers in various industrial segments who use data. As part of this effort, Toshiba has developed a framework for an "Automatic Learning Platform" to allow many customers to use deep learning. Toshiba is assisting the customers, who desire to analyze data more minutely for use in their business, to build an environment for starting deep learning simply and smoothly. Fully utilizing its vast experience and advanced media intelligence technology gained in a variety of fields, Toshiba has launched its unique solutions that deal with problems such as a small volume of data available for learning and relatively unclear images and video for recognition. In conjunction with the provisioning of this framework, Toshiba will provide templates and parameters to deal with customer problems after verifying the results of such solutions in verification experiments. This article introduces advanced systems and schemes developed by Toshiba for deep learning.

Framework for Automatic Learning Platform,
the Fruit of Toshiba's Know-how Gained at Fields

Deep learning is an area of machine learning where a neural network that has learned features in depth provides recognition, analysis and other functions. In a sense, this is an area in which standardized methodologies cannot be established easily. As a result, at present, extensive trial and error by experts and deserved amounts of cost and time are required to introduce deep learning. In many cases, deep learning is started using a publicly available learned model as an open source to save time and effort. In these cases, however, it is not easy for users to realize deep learning that is suitably tailored to their specific purposes. This is because the technologies to be used depending on the content to be analyzed or the types of data such as images, voice and sensor data.

Toshiba has developed a framework that automatically executes a series of processes ranging from computer programming and debugging for operation verification to confirmation of its effects. Users can smoothly start the deep learning by simply choosing from plural choices of network configuration and parameter setting, which have been designed in the past by researchers and experts based on rules of thumb.

This framework maximizes the media intelligence technology nurtured by Toshiba Industrial ICT Solutions Company over a long time and know-how acquired through various industrial region. Toshiba has also developed techniques to solve problems such as "quantity" and "quality" of learning data fully utilizing its in-depth experience, to further enhance the accuracy of deep learning in the industrial region.

Click here to move to the top of this page.

Automated Data Augmentation and Efficient Learning
are the Solutions to "Quantity" Problem of Learning Data

A framework that is delivered to the customers as an automatic deep learning platform. A key technique to solve the problem of "quantity" of learning data is described in this article together with its advanced use cases. Adequate and large volume of quality learning data is a prerequisite in enhancing the accuracy of deep learning. As for images, it is said that tens of thousands of sample images are necessary. It seems quite rare for customers to store such a large volume of images as data. Certainly, there are cases where almost all data required for analysis is retained in the heads of skilled maintenance personnel.

Toshiba has developed a unique technique to supplement the volume of data by automatically augmented learning data through deep learning in the case that the quantity of learning data is not sufficient. Highly effective results of this technique have already been verified. One application example of this technique is the monitoring and inspecting system for electric power infrastructure collaborated with Alpine Electronics, Inc. and Toshiba that will use an industrial drone. This system allows quick detection of places that require an inspection with high accuracy by analyzing images of power transmission lines and other facilities photographed by a camera installed on a drone. Conventionally, inspection work is performed visually by maintenance personnel. When a damage spot is confirmed, an action will be taken and, in some cases, photograph and storage of images of damage spots are not necessarily required. In this conventional method, therefore, the number of stored images of damaged spots will be very limited compared with the number of images required for deep learning. Toshiba has succeeded in producing large quantities of images that are similar to images that show a damaged spot and that do not show a damaged spot by combining and analyzing by deep learning to solve this problem. (See Fig. 1) Toshiba has verified that flaws and trouble spots of power transmission lines can be detected stably with higher accuracy also through deep learning using these augmented images.

Fig. 1 Learning Data is Augmented by Deep Learning

Toshiba developed a framework that enables a computer to learn a large quantity of sample images efficiently from various viewpoints. Toshiba has built a system that automatically manages learning models created through diverse learning by correlating them to their inference indicators, to learning data and to other items. As a result, an environment for deep learning to supplement a shortage of learning data has been established.

Click here to move to the top of this page.

Movements and Motions of Rugby Players are Detected
with High Accuracy and are Analyzed in Time Series

Shinichi Kashimoto

Active effort is also made to improve the "quality" of data.
High accuracy of deep learning through recognition of images, video and voice at a high level will allow various types of forecast and analysis by combining other information and technologies. One example is the self-driving system, which has become a topic item, for avoidance of danger by forecasting the movements and motions of automobiles and people, in addition to the function of finding an obstacle. If the movements and motions of people, operation time and other parameters in various processes of factories could be analyzed with high accuracy, it will be possible to greatly enhance quality and productivity in the manufacturing industry.

With a view to applying this notion to the industrial region, Toshiba is active in achieving a higher accuracy of image recognition through deep learning using sporting scenes in which players complicatedly move and act swiftly. Verification experiments of this effort have been undertaken. In this experiment, approximately 1,000 image for learning data are extracted from game video recorded by the staff of the rugby team Toshiba BRAVE LUPUS, in addition to learning image data for general object recognition. Using the deep learning, players could be recognized from the play video with higher accuracy than that achieved by the conventional technology. Movements of players and ball are tracked in time series to infer positions of the players and ball. By adding the recognition of video voice such as whistling by referees, play scenes including "tackles" and "scrums" can now be recognized. At present, a very large amount of time and work is needed to analyze play scenes such as an analysis of the timing for players to start moving, team formation and factors that produce specific play results. Toshiba plans to demonstrate that this play analysis can be performed quickly and timely by analyzing rugby images through deep learning. (See Fig. 2)

Fig. 2 Video Data Analysis of Rugby Play

The technology to accurately recognize the movements and motions of players and the ball in a complex play, which is a characteristic of rugby, is enabled through precise learning. Toshiba plans to apply this technology to a management of traffic lines, to a verification of work performed, to a measurement of operation time and to other purposes, in order to improve the quality and productivity in the manufacturing industry.

Click here to move to the top of this page.

Support Customers Everlastingly
even after Launch of Deep Learning

Accuracy of deep learning is expected to surpass that of humans in the future. Applications of deep learning to various scenes of the industry and society are eagerly awaited. Toshiba is undertaking various projects to accelerate the advances of deep learning. However, Toshiba believes it is not necessary to change the means of data analysis required for customer businesses to deep learning entirely. Rather, Toshiba believes that deep learning could practically solve problems that customers have not enough cost and time to spend on analysts and experts and could effectively reinforce conventional analysis where accuracy is still low. Toshiba strongly wishes to make deep learning to be a mean that can be started without difficulty to solve problems practically and effectively. The framework for automatic learning platform is a tool for this purpose. By continuing to add the knowledge and know-how which Toshiba has uniquely acquired through verification experiments and other efforts to this tool, the accuracy of the analysis will increase steadily, making it possible to achieve a reform and discovery, which the industry and society have not experienced before. This will enable to solve the problems that have troubled customers over a long time. Toshiba will continue to pursue the ideal image of deep learning that is truly valuable while cherishing a variety of dreams.

* The corporate names, organization names, job titles and other names and titles appearing in this article are those as of January 2017.

Related articlesVol.20: Artificial Intelligence Discovering Values of IoT Data -Analytics towards Deep Learning-