Info
- Title: Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
- Task: Image Classification
- Author: C. Szegedy, S. Ioffe, V. Vanhoucke, and A. Alemi
- Arxiv: 1602.07261
- Date: Feb. 2016
Highlights & Drawbacks
The article mentioned in the experimental part that the Inception network can be upgraded to the SOTA level without the help of the Skip-connection structure, but adding Skip-connection can effectively increase the training speed.
Motivation & Design
In 2015, ResNet became the most dazzling convolutional network structure of the year, and the structure of skip-connection became an option to avoid. The Inception series also updated its structure with reference to ResNet. At the same time, the fourth generation and the combination with ResNet were introduced: Inception-v4 and Inception-ResNet.
Inception-v4 Architecture
NetScope Visualization and source code: awesome_cnn
Inception-ResNet(v2) Architecture
NetScope Visualization and source code: awesome_cnn
