https://towardsdatascience.com/mobilenetv2-inverted-residuals-and-linear-bottlenecks-8a4362f4ffd5

In April 2017 a group of researchers from Google published a paper which introduced a neural network architecture that was optimized for mobile devices. They strived for a model that delivered high accuracy while keeping the parameters and mathematical operations as low as possible. This was much needed in order to bring deep neural networks to smartphones.

The architecture dubbed MobileNet revolves around the idea of using depthwise separable convolutions, which consist of a depthwise and a pointwise convolution after one another. If you’re a little fuzzy on the details of this operation feel free to check out my other article that explains this concept in detail. MobileNetV2 extends its predecessor with 2 main ideas.