====== 1 x 1 convolution ======

Changes number of channels in volumes.

====== Inception network ======

Instead of choosing filter size, use em all.

28x28x192

Try 1x1 => 28x28x64
Try 3x3 => 28x28x128
Try 5x5 => 28x28x32
Try MaxPool => 28x28x32

Output: 28 x 28 x 256

Problem: Computational cost

28x28x192 => Conv 5x5, same 32 => 28x28x32

28x28x32 x 5x5x192 = 120 M

Using 1x1 convolution as bottleneck layer:

28x28x192 => Conv 1x1, 16 => 28x28x16 => Conv 5x5, 32 => 28x28x32

  * 28x28x16 x 1x1x192 = 2,4 M
  * 28x28x32 x 5x5x16 = 10 M
  * Total: 12,4 M multiplications (1/10)