NEC develops deep learning technology to improve recognition accuracy
Via NEC News room
Dec 12, 2017
Tokyo, December 12, 2017 - NEC Corporation (NEC; TSE: 6701) today announced it has developed automatic optimization technology for deep learning in order to facilitate improvements in recognition accuracy.
In recent years, there have been tremendous advances in deep learning, which is now contributing to image recognition, speech recognition, and a wide range of other fields. Deep learning enables higher levels of recognition accuracy by capitalizing on the deeply layered structures of artificial neural networks in order to learn from prepared data.
If systems become excessively familiar with data however, they become unable to accurately recognize data that they have not learned. This is known as "overtraining," and results in degradation of recognition accuracy when dealing with data that was not used in the learning process. To prevent overtraining, "regularization" technology is commonly used, which regulates the extent of learning to prevent it from reaching an excessive degree.
"This technology predicts the progress of learning at every layer based on the structure of an artificial neural network, and enables regularization to be automatically configured accordingly," said Akio Yamada, General Manager of NEC's Data Science Research Laboratories. "This means that learning is optimized across the entire network, making it possible to improve recognition accuracy, such as reducing recognition errors by around 20% when compared to conventional systems."
"This technology is expected to improve recognition accuracy for image and speech recognition, and a whole host of other fields in which deep learning is used," continues Yamada. "It will be able to improve the accuracy of facial recognition and behavior analysis for purposes that include video surveillance, for instance, or to increase the efficiency of inspections of infrastructure, or enable the automatic detection of system failures, accidents or disasters."
Automatic configuration of regularization at each layer of an artificial neural network