NEC has developed a deep learning automatic optimization technology, easier to improve the recognition accuracy

Recently, NEC announced that it has developed a deep learning auto-optimization technology that can more easily improve the recognition accuracy.

In the past, deep learning was difficult to adjust based on the neural network structure (Note 1), so that it was impossible to perform optimal learning throughout the entire network, thus failing to fully realize its identification. The developed technology can be automatically optimized based on its structure The progress of neural network learning makes it easier to identify more accurately than ever before.

The emergence of this technology makes it possible to further improve the recognition accuracy in various fields such as image recognition and voice recognition, etc. For example, recognition accuracy of video surveillance such as face recognition and behavior analysis, and infrastructure Check the efficiency of work to achieve automatic detection of disasters, accidents and disasters.

First, the background

In recent years, the research on deep learning has made leaps and bounds, and has been applied in a wide range of fields such as image recognition and voice recognition, etc. Deep learning uses a neural network with deep structure to learn data prepared in advance to achieve high precision. However, if the data is overly learned, there is a phenomenon of "over-learning (Note 2)", in which only the learned data can be identified with high precision, and the recognition accuracy of data not used for learning is reduced. The occurrence of this situation requires the use of regularization (Note 3) technology to adjust.

Due to the complicated and changeable learning process of neural networks, the same regularization technique can only be used for the entire network in the past, resulting in problems such as excessive learning at all levels of the network and some learning stagnation, which makes it difficult to fully In addition, since it is extremely difficult to manually adjust the learning progress of all levels, the demand for automatically adjusting the learning progressively by layer is very high.

The technology developed this time is based on the structure of the neural network, which predicts the progress of learning layer by layer and automatically configures regularization techniques that are suitable for the progress of all layers. With this technique, learning is optimized throughout the network and the recognition error rate can be reduced by about 20%, improve the recognition accuracy.

'Figure' layer of neural network regularization technology automatically set the schematic

Second, the advantages of new technologies

1, according to the neural network structure of automatic learning optimization

Based on the neural network structure, we predict the learning progress of each layer, and automatically set the regularization suitable for the progress of each layer layer by layer. Thus, the learning progress of the entire network is optimized to solve the problems of excessive learning and learning Stagnation.In recognition experiments using handwritten digital data of this technique, the recognition error rate has been reduced by about 20%, and the recognition accuracy has been significantly improved.

The change in the recognition error rate of 'graph' against the amount of learning data

2, with the same amount of calculations in the past, easy to achieve high accuracy

This technique is implemented once only prior to learning a neural network, making it easy to achieve high accuracy with the same amount of computational learning as ever.

NEC Group is committed to advancing social solutions worldwide to provide secure, secure, efficient and fair social values, to integrate advanced ICT technologies with knowledge, and to contribute to a brighter, more colorful and efficient society .

(Note 1) Neural Network: A neural network composed of artificial nerve cells (neurons).

(Note 2) Over-learning: over-learning given data, but not learning data

Recognition of the phenomenon of decreased accuracy.

(Note 3) Regularization: Suppression of over-learning by constraining the complexity of the model.

The most cool new media technology information

Sweep won the prize