NEC develops deep learning automatic optimization technology

Recently, NEC announced the development of "deep learning automatic optimization technology" to improve identification accuracy. In the past, when deep learning was performed, it was difficult to adjust the learning method according to the structure of the neural network. Therefore, it was impossible to optimize the entire network learning, and thus the original identification performance could not be fully utilized. The technology developed by NEC, with the progress of neural network learning, is automatically optimized according to its structure, and can easily achieve higher recognition accuracy than in the past.

Using this technology, the accuracy of recognition can be further improved in various fields such as image recognition and sound recognition using deep learning techniques. For example, it improves the accuracy of image monitoring such as face recognition and behavior analysis, improves efficiency when performing maintenance inspections at infrastructures, etc., and is expected to automatically detect faults, accidents, or disasters.

In recent years, deep learning has made great progress. Based on image recognition and sound recognition, it is widely used in different fields. The so-called deep learning is to use a neural network with multi-layer structure to let the computer learn the data prepared in advance, and thus improve the identification accuracy. However, if the computer over-learns the data, there will be an "over-training" phenomenon, that is, only the learned data will have a higher recognition accuracy, and the accuracy will be reduced when identifying the data that has never been learned. In order to avoid this, the "normalization" method is usually used to adjust the process of deep learning.

The learning process of a neural network will produce complex changes in response to the structure, so in the past, the same formalization can only be performed on the entire neural network. As a result, among the various layers of the neural network, some have excessive training and some cannot learn smoothly, so it is difficult to fully utilize the original identification performance. In addition, because it is extremely difficult to manually adjust the learning progress of each layer one by one, the demand for automation adjustment on the market is also quite high.

The technology developed by NEC is based on the structure of the neural network, predicting the learning progress of each layer, and automatically setting the normalization layer by layer according to the learning progress of each layer. Through such a technology, the learning of the entire neural network can be optimized, and the recognition error rate can be reduced by 20% compared with the conventional method, and the identification accuracy is improved.

The advantages of new technology

According to the structure of the neural network, the learning situation is automatically optimized: according to the structure of the neural network, the learning progress of each layer is predicted, and the normalization is automatically set layer by layer according to the learning progress of each layer. Through such a technology, it is possible to optimize the learning of the entire neural network, and also solve the problem of excessive training and inability to learn smoothly in the past. Not only that, the use of this technology

The calculation is the same as in the past, and it is easy to achieve high precision: before the deep learning of the neural network, it is only necessary to run the technology once, even if the calculation amount of learning is the same as in the past, it can easily achieve high precision. The NEC Group is committed to promoting the "social solutions business" globally to provide security. Peace of mind. effectiveness. A fair social value, combined with advanced ICT technology and knowledge, to achieve a brighter, more abundant, more efficient and quintessential society.

Voltage Regulator

PC-TM series Relay Control Voltage Regulator has the low energy consumption,the over voltage protection,the low voltage protection,the over-current protection,the over-loading protection,the over-temperature protection and so on.It boasts for many kinds of protections,the collection energy conservation and the environmental protection ect.This is a brand-new concept product which possess many new technologies!This series products simultaneously ha applied for many technical monopolies

We already applied many kinds of this products patent, and the technical patent NO: 200720036394.1 and Appearance paten NO: 200730025909.3

2. Use for equipment:

Computer

Test equipment

Light system

Safe alarm system

Ray equipment

Medical equipment

Copy machine

Stereo equipment

Numerical control machine tools

Industrial automation equipment