Google’s new neural network is much more accurate

Google’s new neural network is much more accurate

Researchers applying the same ladder method are suitable for tasks such as object and person recognition, but to improve their accuracy requires boring and fine-tuning. That’s why Google AI researchers are investigating new models that “scaling” CNN in a “more structured” way.

The results of their work are published in the article “EfficientNet: Reviewing the Neural Networks Scale” on the Arxiv.org portal. Co-authors argue that the EfficientNets family of artificial intelligence systems exceeds the accuracy of standard CNN and increases the performance of neural networks up to 10 times.

Unlike the standard CNN scaling approach, the Google AI team offers a straightforward scaling method for all neuron network parameters. “The usual practice in scaling models is to arbitrarily increase the depth or width of CNN, and to use a larger picture of the input image for training and evaluation,” notes Eng.

“Unlike traditional approaches that randomly scans network parameters such as width, depth and input resolution, our method scales evenly every dimension with a fixed set of zooming factors.”

In order to further increase productivity, researchers are offering the use of a new network base network – the Mobile Inverted Convolution in Closer Places (MBConv), which serves as the basis for EfficientNets family models. In EfficientNets, it shows greater accuracy and greater efficiency than existing CNNs and decreases in the order of parameter and computer resources.

One of the models, EfficientNet-B7, showed 8.4 times smaller size and 6.1 times more performance than the famous CNN Gpipe, reaching 84.4% and 97.1% (Top-1 and Top 5) accuracy when test ImageNet. Compared to the popular CNN ResNet-50, the other EfficientNet-EfficientNet-B4 model using similar resources showed a 82.6% accuracy compared to 76.3% for ResNet-50.

EfficientNets models work well with other data sets, achieving high precision in five out of eight tests, including CIFAR-100 (91.7% accuracy) and Flowers (98.8%).

“By providing significant improvements in the performance of neural models, we expect EfficientNet to potentially serve as a new basis for future computing tasks,” wrote Tang and Lee. Source code and Google TPU cloud trick training tricks are available for free at Github.

Google’s new neural network is much more accurate