New data processing engine makes deep neural networks smarter

“Feature normalization is an important element of training deep neural networks, and attention to

functions is equally important to helpnetworks to highlight which features extracted from raw data are most important for completing tasks,” explains Tianfu Wu, assistant professor of electrical and computer engineering at NC State. “But mostly they were processed separately. We found that combining them made them more efficient and effective.”

To test your AN module,researchers have connected it to the four most widely used neural network architectures: ResNets, DenseNets, MobileNetsV2, and AOGNets. They then tested the networks against two industry standard metrics: ImageNet-1000 classification test and object detection and instance segmentation 2017 MS-COCO test.

“We found that AN improved performancefor all four architectures in both tests, ”Wu said. “For example, the Top-1 accuracy in ImageNet-1000 improved by 0.5-2.7%. The average precision (AP) accuracy increased to 1.8% for the bounding box and 2.2% for the semantic mask in MS-COCO. Another advantage of AN is that it facilitates better transfer of learning between different domains. For example, from image classification in ImageNet to object detection and semantic segmentation in MS-COCO. This is illustrated by the performance improvement in the MS-COCO benchmark, which was obtained by fine-tuning the deep neural networks previously trained by ImageNet at MS-COCO. "

“We have released the source code and hope that our AN will lead to better integrative design for deep neural networks,” the scientists conclude.

Read also

The Doomsday glacier turned out to be more dangerous than scientists thought. We tell the main thing

Two pieces of evidence of extraterrestrial life emerged at once. One on Venus, the other - no one knows where

It turned out that the moons of Uranus are more like planets