MobileNetv2 linear bottleneck layers

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

Hi all,

In Week 2 Assignment 4 (of course 4), it says:
“Each block consists of an inverted residual structure with a bottleneck at each end. These bottlenecks encode the intermediate inputs and outputs in a low dimensional space, and prevent non-linearities from destroying important information.”

I’m not sure about why the latter part (bolded) is true - how do bottleneck layers work to accomplish this? In the paper, they mention “linear bottleneck layers” - which I guess applies some linear transformation? [1]

[1]. https://arxiv.org/pdf/1801.04381.pdf

2

Sorry, this question was missed or skipped when it was originally asked. Just happened to find it while searching for “bottleneck”. Here’s a more recent thread with a similar starting question and maybe not a definitive answer, but better than nothing. :nerd_face: