Experience in Network Structure

Improvement for Feature Expression

inverted residual with linear bottleneck

ReLU6 、 ReLU、PReLU

Acceleration and Compression

Depthwise Separable Convolution

computational cost reduce from

to

Group Convolution

from 35x35x320 to 17x17x640:

  1. conv+pooling or pooling + conv
  2. conv with stride 2

  3. remain to study which is faster between 2 and 3.

Inspired Architecture

Insight

Manifold of interest should lie in a low-dimensional subspace of the higher-dimensional activation space (non-linearity destroys information in low-dimensional space.)