# Rethinking the Inception Architecture for Computer Vision

**Authors:** Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, Zbigniew Wojna  
**arXiv:** [1512.00567](https://arxiv.org/abs/1512.00567)  
**Date:** 2015-12-02  
**Summary page:** https://www.summarizepaper.com/en/paper/1512.00567v3/

## Key points

- Trend towards very deep networks since 2014 with impressive performance on various benchmarks
- Importance of computational efficiency and low parameter count in applications like mobile vision and big-data scenarios
- Proposal of innovative methods such as factorized convolutions and rigorous regularization techniques to efficiently scale up networks
- Significant progress demonstrated over existing models on the ILSVRC 2012 classification challenge validation set
- Achievement of top-1 error rate of 21.2% and top-5 error rate of 5.6% using a network with specific computational cost and parameter count
- Impressive results through ensemble modeling and multi-crop evaluation, showcasing effectiveness in pushing boundaries while maintaining efficiency
- Emphasis on optimizing network architectures for improved performance in real-world applications with resource constraints

## Summary

The paper "Rethinking the Inception Architecture for Computer Vision" by Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna explores advancements in convolutional networks that have transformed computer vision solutions. The authors highlight the trend towards very deep networks since 2014 and their impressive performance on various benchmarks. While larger models and increased computational costs usually lead to better results with sufficient labeled data for training, the importance of computational efficiency and low parameter count cannot be ignored in applications like mobile vision and big-data scenarios. To address this issue, the authors propose innovative methods such as factorized convolutions and rigorous regularization techniques to efficiently scale up networks. By evaluating their approaches on the ILSVRC 2012 classification challenge validation set, they demonstrate significant progress over existing models. Specifically, they achieve a top-1 error rate of 21.2% and a top-5 error rate of 5.6% using a network with a computational cost of 5 billion multiply-adds per inference and less than 25 million parameters. Additionally, through an ensemble of four models and multi-crop evaluation, they report an impressive top-5 error rate of 3.5% on the validation set (3.6% on the test set) and a top-1 error rate of 17.3% on the validation set. This showcases the effectiveness of their approach in pushing the boundaries of computer vision capabilities while maintaining computational efficiency. The study emphasizes the significance of optimizing network architectures for improved performance in real-world applications where resource constraints are prevalent.