EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

By M. Tan et al.

Published on June 10, 2019

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Related Work
3. Compound Model Scaling
4. EfficientNet Architecture

Summary

EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks is a paper that systematically studies model scaling for Convolutional Neural Networks (ConvNets) and proposes a new scaling method that balances network depth, width, and resolution to achieve better performance. The paper introduces a compound scaling method that uniformly scales all dimensions of depth/width/resolution using a fixed ratio. It demonstrates the effectiveness of this method on scaling up MobileNets and ResNet, leading to the development of a new family of models called EfficientNets. These EfficientNets achieve state-of-the-art accuracy on ImageNet and other transfer learning datasets while being more efficient in terms of parameters and inference speed compared to existing ConvNets. The paper also discusses the importance of balancing network width, depth, and resolution during ConvNet scaling for better accuracy and efficiency. Furthermore, it presents the EfficientNet architecture, developed through neural architecture search, as an efficient baseline network for demonstrating the effectiveness of the proposed scaling method.