Fujitsu Deep Neural Networks (DNN)

Deep Neural Networks (DNN) is technology of deep training.

On May 25, 2017 the Fujitsu company announced creation of the mechanism of distribution of memory for "deep neural networks" (Deep Neural Networks, DNN).

The created technology can be used:

• in medical practice for analytics - detection of a diabetic retinopathy,
• for the analysis and classification of satellite pictures,
• for natural languag processing,
• processings of large volumes of data on the basis of graphs (including devices on the basis of Internet of Things),
• in financial transactions,
• as a part of social networks, etc.

Representation of memory relocation, (2017)

The DNN networks used in different areas of the sphere of artificial intelligence, including recognition and classification of the speech and objects require the large volume of computing resources. It creates a heavy load on the operating computing infrastructures. Within development for deep training model parallelism is used in problems of automatic load distribution for memory of DNN networks. As a result, possibilities of the infrastructures for data processing processed by applications of artificial intelligence considerably extend without the need for additional investments.

In recent years we observe emergence of developments in which hardware accelerators for support of large volume of calculations of DNN networks are used. Permanent increase in expenses on calculations in DNN networks represents a serious problem, especially when the size of model of DNN network increases to such size that it cannot be located in memory of one accelerator. At problem solving, connected with artificial intelligence, wider and deep neural networks and also more accurate classification of categories are required. Our development allows to solve directly this problem, distributing memory requirements of DNN networks on several computers. Using our technology it is possible to increase the size of neural networks to several computers for creation of more exact and large-scale models of DNN networks. Tsuneo Nakata, chief executive officer of Fujitsu Laboratories of Europe

The technology allows to distribute memory by means of conversion disorderly of organized neural networks in equivalent networks where separate or all levels are replaced with a set of smaller subtotals. These subtotals are created so that to become a complete analog of original levels, but differ in more high efficiency of calculations. As original and other levels come from the same profile, learning process of distributed networks of DNN converges with original network DNN without any additional expenses.

Robotics