# Network In Network

**Authors:** Min Lin, Qiang Chen, Shuicheng Yan  
**arXiv:** [1312.4400](https://arxiv.org/abs/1312.4400)  
**Date:** 2013-12-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1312.4400v3/

## Key points

- "Network In Network" proposes a new deep network structure to improve model discriminability for local patches within the receptive field.
- Conventional convolutional layers use linear filters followed by nonlinear activation functions to scan input.
- Building micro neural networks with more complex structures can better abstract data within the receptive field.
- Micro neural networks are instantiated with multilayer perceptrons which are potent function approximators.
- Feature maps are obtained by sliding these micro networks over input in a similar manner as CNNs and then fed into the next layer.
- Deep NIN is implemented by stacking multiple of these structures.
- With enhanced local modeling via the micro network, global average pooling over feature maps in the classification layer is utilized instead of traditional fully connected layers; this approach is easier to interpret and less prone to overfitting.
- The authors demonstrate state-of-the-art classification performance with NIN on CIFAR-10 and CIFAR-100 datasets and reasonable performance on SVHN and MNIST datasets.
- The paper's authors are Min Lin, Qiang Chen, and Shuicheng Yan.

## Summary

The paper "Network In Network" proposes a novel deep network structure to improve model discriminability for local patches within the receptive field. Conventional convolutional layers use linear filters followed by nonlinear activation functions to scan input; however, the authors argue that building micro neural networks with more complex structures can better abstract data within the receptive field. These micro neural networks are instantiated with multilayer perceptrons which are potent function approximators. Feature maps are obtained by sliding these micro networks over input in a similar manner as CNNs and then fed into the next layer. Deep NIN is implemented by stacking multiple of these structures. With enhanced local modeling via the micro network, global average pooling over feature maps in the classification layer is utilized instead of traditional fully connected layers; this approach is easier to interpret and less prone to overfitting. The authors demonstrate state-of-the-art classification performance with NIN on CIFAR-10 and CIFAR-100 datasets and reasonable performance on SVHN and MNIST datasets. The paper's authors are Min Lin, Qiang Chen, and Shuicheng Yan.