Enhancing ResNet to ResNeXt for image classification
What is ResNeXt
After ResNet, I would like to talk about ResNeXt which is designed for image classification. To improve accuracy, we may stack more layers or use widen layers but the problem is more parameters are introduced as well. Eventually, more computational power is required. Therefore, Xie et al., propose a new architecture that improving accuracy while reducing network complexity and the number of parameters.
This story will discuss Aggregated Residual Transformations for Deep Neural Networks (Xie et al., 2016) and the following are will be covered:
- Architecture
- Experiment
Architecture
Inherited from ResNet, VGG, and Inception, ResNeXt includes shortcuts from the previous block to next block, stacking layers and adapting split-transform-merge strategy.
- ResNet: Introducing a shortcut from the previous layer to the next layer
- VGG: Leveraging repeating layers to build a deep architecture model
- Inception: Following split-transform-merge practise to split the input to multiple blocks and merging blocks later on.
- ResNeXt: The principle is stacking the same topology blocks. Within the residual block, hyper-parameters…