Very Deep Convolutional Networks for Large-Scale Image Recognition

Authors: Simonyan, K., & Zisserman, A. Year: 2014 Venue: ICLR ArXiv: https://arxiv.org/abs/1409.1556

Summary

VGGNet demonstrates that network depth is critical for accuracy. Using small 3x3 filters throughout, VGGNet achieves state-of-the-art on ImageNet and became the foundation for many modern architectures.

Key Concepts

• Deep architecture (16-19 layers)
• Small 3x3 convolutional filters
• Stacking layers instead of large filters
• Systematic architecture exploration
• Transfer learning through pre-training

Impact

VGGNet showed depth matters and influenced architecture design for a decade. VGG features became the standard for transfer learning tasks.

Related Papers

• AlexNet: ImageNet Classification
• ResNet: Deep Residual Learning

---

Citation:

@inproceedings{simonyan2014very,
  title={Very deep convolutional networks for large-scale image recognition},
  author={Simonyan, Karen and Zisserman, Andrew},
  booktitle={International Conference on Learning Representations},
  year={2014}
}