EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Authors: Tan, M., & Le, Q. V. Year: 2019 Venue: ICML ArXiv: https://arxiv.org/abs/1905.11946

Summary

EfficientNet proposes a principled approach to scaling CNNs across depth, width, and resolution. It achieves better accuracy-efficiency tradeoff than previous models, becoming standard for efficient vision.

Key Concepts

• Compound scaling method
• Systematic architecture search (NAS)
• Depth, width, resolution balance
• State-of-the-art efficiency
• Transferable scaling principles

Impact

EfficientNet set new standards for model scaling and efficient deep learning. EfficientNet-B0 to B7 family covers the entire efficiency spectrum.

Related Papers

• MobileNet: Efficient Convolutional Networks
• ShuffleNet: Channel Shuffle
• Neural Architecture Search

---

Citation:

@inproceedings{tan2019efficientnet,
  title={EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks},
  author={Tan, Mingxing and Le, Quoc V},
  booktitle={International Conference on Machine Learning},
  year={2019}
}