Deep Learning - Bottleneck Structure (Bottleneck)

1. Introduction
Paper: One of the core contents of Deep Residual Learning for Image Recognition
ResNet is "Deeper Bottleneck Architectures" (DBA for short). In a word, bottleneck is a special residual structure.

The original picture in the Resnet paper is as above (ie Bottleneck V1), the left picture is the ordinary residual structure, and the right picture is the bottleneck structure. Specifically, the input and output channel_num of the block are the same (256 in the upper right picture and 64 in the left picture),

And the channel_num in the block structure (64 in the upper right picture) is smaller than the input/output channel_num (256), which is very vivid. That is to say, reduce the dimension first and then increase the dimension.

2. Function
After switching to bottleneck design, the parameters of the network are reduced a lot, the depth is also deepened, and the training is relatively easy.

Reference:
Deep Learning Foundation – Bottleneck (Bottleneck) Architectures