# XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks

**Authors:** Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, Ali Farhadi  
**arXiv:** [1603.05279](https://arxiv.org/abs/1603.05279)  
**Date:** 2016-03-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1603.05279v4/

## Key points

- Authors propose two efficient approximations to standard convolutional neural networks (CNNs): Binary-Weight-Networks and XNOR-Networks
- Binary-Weight-Networks approximate filters in CNNs with binary values, resulting in a 32x reduction in memory usage without sacrificing accuracy
- XNOR-Networks represent both filters and input as binary values, enabling primarily binary operations and leading to a 58x increase in computational speed and an additional 32x reduction in memory usage compared to traditional CNNs
- Binary-Weight-Network version of AlexNet achieves only a 2.9% lower classification accuracy compared to full-precision AlexNet on ImageNet task (measured using top-1 accuracy)
- XNOR-Nets outperform recent network binarization methods such as BinaryConnect and BinaryNets by large margins on ImageNet, achieving more than a 16% improvement in top-1 accuracy
- Combination of reduced memory usage and increased computational speed makes XNOR-Nets highly practical for real-time applications on devices with limited computing resources
- Proposed binary networks are simple, accurate, efficient, and capable of handling complex visual tasks.

## Summary

In their paper titled "XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks," authors Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi propose two efficient approximations to standard convolutional neural networks (CNNs): Binary-Weight-Networks and XNOR-Networks. These approximations aim to reduce memory usage and increase computational efficiency. In Binary-Weight-Networks, the filters in CNNs are approximated with binary values, resulting in a 32x reduction in memory usage without sacrificing accuracy. XNOR-Networks take the approximation further by representing both the filters and the input to convolutional layers as binary values. This approach enables convolutions to be performed using primarily binary operations, leading to a 58x increase in computational speed and an additional 32x reduction in memory usage compared to traditional CNNs. The authors evaluate their approach on the challenging ImageNet classification task and find that a Binary-Weight-Network version of AlexNet achieves a classification accuracy that is only 2.9% lower than the full-precision AlexNet (measured using top-1 accuracy). Furthermore, when compared to recent network binarization methods such as BinaryConnect and BinaryNets, XNOR-Nets outperform them by large margins on ImageNet, achieving more than a 16% improvement in top-1 accuracy. The combination of reduced memory usage and increased computational speed makes XNOR-Nets highly practical for real-time applications such as deploying high performance CNN models on devices with limited computing resources. Overall, the proposed binary networks are simple, accurate, efficient, and capable of handling complex visual tasks.