# EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

**Authors:** Mingxing Tan, Quoc V. Le  
**arXiv:** [1905.11946](https://arxiv.org/abs/1905.11946)  
**Date:** 2019-05-28  
**Summary page:** https://www.summarizepaper.com/en/paper/1905.11946v5/

## Key points

- Authors explore model scaling in Convolutional Neural Networks (ConvNets)
- Current practice is to scale up for better accuracy if more resources become available
- Authors propose a new scaling method that uniformly scales all dimensions of depth/width/resolution using a compound coefficient
- This method is highly effective on MobileNets and ResNet models
- Authors use neural architecture search to design a new baseline network which they scale up to obtain EfficientNets family of models
- EfficientNets achieve much better accuracy and efficiency than previous ConvNets
- EfficientNet-B7 achieves state-of-the-art 84.3% top-1 accuracy on ImageNet while being 8.4x smaller and 6.1x faster on inference than the best existing ConvNet
- EfficientNets transfer well across datasets and achieve state-of-the-art accuracy on CIFAR-100 (91.7%), Flowers (98.8%), and three other transfer learning datasets with an order of magnitude fewer parameters than previous models

## Summary

In their paper "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks," authors Mingxing Tan and Quoc V. Le explore the concept of model scaling in Convolutional Neural Networks (ConvNets). They note that ConvNets are often developed with a fixed resource budget and then scaled up for better accuracy if more resources become available. However, the authors argue that carefully balancing network depth, width, and resolution can lead to even better performance. To this end, they propose a new scaling method that uniformly scales all dimensions of depth/width/resolution using a compound coefficient. This method is demonstrated to be highly effective on MobileNets and ResNet models. The authors take things further by using neural architecture search to design a new baseline network which they scale up to obtain a family of models called EfficientNets. These models achieve much better accuracy and efficiency than previous ConvNets; their EfficientNet-B7 achieves state-of-the-art 84.3% top-1 accuracy on ImageNet while being 8.4x smaller and 6.1x faster on inference than the best existing ConvNet. Furthermore, the EfficientNets transfer well across datasets and achieve state-of-the-art accuracy on CIFAR-100 (91.7%), Flowers (98.8%), and three other transfer learning datasets with an order of magnitude fewer parameters than previous models. Overall, the authors' work highlights the importance of careful model scaling in ConvNets and presents an effective new method for achieving improved performance while maintaining efficiency. Their EfficientNet models demonstrate significant improvements over previous state-of-the-art models across multiple datasets while requiring fewer parameters.