# Learning Deep Features for Discriminative Localization

**Authors:** Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, Antonio Torralba  
**arXiv:** [1512.04150](https://arxiv.org/abs/1512.04150)  
**Date:** 2015-12-14  
**Summary page:** https://www.summarizepaper.com/en/paper/1512.04150v1/

## Key points

- The paper discusses the use of global average pooling layer in convolutional neural networks (CNNs) for remarkable localization ability.
- Global average pooling was initially proposed as a regularization technique but is found to build a generic localizable deep representation.
- The network achieves a top-5 error of 37.1% for object localization on ILSVRC 2014, close to fully supervised CNN approach with 34.2% top-5 error.
- Global average pooling enables CNNs to achieve remarkable localization ability and produce generic deep representations.
- The approach is versatile and effective, capable of localizing discriminative image regions across different tasks without specific training.

## Summary

The paper "Learning Deep Features for Discriminative Localization" by Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba revisits the global average pooling layer proposed in a previous work. The authors shed light on how this layer enables convolutional neural networks (CNNs) to have remarkable localization ability despite being trained on image-level labels. Initially proposed as a means for regularizing training, global average pooling is found to actually build a generic localizable deep representation that can be applied to various tasks. Despite its apparent simplicity, this approach achieves impressive results. The authors demonstrate that their network achieves a top-5 error of 37.1% for object localization on ILSVRC 2014, which is remarkably close to the 34.2% top-5 error achieved by a fully supervised CNN approach. This highlights the effectiveness of global average pooling in enabling CNNs to achieve remarkable localization ability and build generic deep representations that can be applied to diverse tasks. The versatility of global average pooling is emphasized by the authors as their network is capable of localizing discriminative image regions across different tasks, even though it was not specifically trained for them. This showcases the power of this approach in producing effective results for various applications. are utilized in this paper to enable , which is achieved through . The key component responsible for this success is , which allows for regularization during training while also building a generic deep representation that can be applied to diverse tasks with impressive results. The versatility and effectiveness of this approach are highlighted throughout the paper, showcasing the power of global average pooling in enabling CNNs to achieve remarkable localization ability and produce generic deep representations.