Google's new neural network is much more accurate and faster than popular counterparts

Convolutional neural networks (CNNs), inspired by biological processes in the human visual cortex, are well suited for tasks such as object and face recognition, but improving their accuracy requires tedious and fine tuning. That's why scientists at Google AI Research are exploring new models that "scale" CNNs in a "more structured" way. They published the result of their work in article "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks", hosted by the scientific portal Arxiv.org, as well as in ARTICLES on your blog. The co-authors claim that the family of artificial intelligence systems, called EfficientNets, exceeds the accuracy of standard CNNs and improves the efficiency of a neural network by up to 10 times.

“A common practice for scaling models is to arbitrarily increase the depth or width of the CNN, as well as use a higher resolution of the input image for training and evaluation,” write in-house software engineer Mingxing Tan and lead scientist at Google AI Quoc Li (Quoc V .le). "Unlike traditional approaches that arbitrarily scale network parameters such as width, depth, and incoming resolution, our method uniformly scales each dimension with a fixed set of scaling factors."

To further improve performance, the researchers advocate the use of a new core network, the mobile inverted bottleneck convolution (MBConv), which serves as the basis for the EfficientNets family of models.

In tests, EfficientNets has demonstrated both higher accuracy and better efficiency than existing CNNs, reducing the requirement for parameter size and computational resources by an order of magnitude. One of the models, EfficientNet-B7, demonstrated 8,4 times smaller size and 6,1 times better performance than the well-known CNN Gpipe, and also achieved 84,4% and 97,1% accuracy (Top-1 and Top- 5 result) in testing on the ImageNet set. Compared to the popular CNN ResNet-50, another EfficientNet model, EfficientNet-B4, using similar resources, showed an accuracy of 82,6% versus 76,3% for ResNet-50.

The EfficientNets models performed well on other datasets, achieving high accuracy in five out of eight tests, including the CIFAR-100 (91,7% accuracy) and Flowers (98,8%).

“By delivering significant improvements in the performance of neural models, we expect EfficientNets to potentially serve as a new foundation for future computer vision challenges,” Tan and Li write.

The source code and tutorial scripts for Google Cloud Tensor Processing Units (TPUs) are freely available at Github.

Source: 3dnews.ru