In this section, we'll discuss a lightweight CNN model called MobileNet (MobileNetV2: Inverted Residuals and Linear Bottlenecks, https://arxiv.org/abs/1801.04381). We'll focus on the second revision of this idea (MobileNetV1 was introduced in MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications, https://arxiv.org/abs/1704.04861).
MobileNet is aimed at devices with limited memory and computing power, such as mobile phones (the name kind of gives it away). To reduce its footprint, the network uses DSC, linear bottlenecks, and inverted residuals.
We are already familiar with DSC, so let's discuss the other two:
- Linear bottlenecks: To understand this concept, we'll quote the paper: