Maybe we can reduce the amount of computation needed by applying 1x1 convolutions first? → Inception
Figure 2
Xception
Hypothesis: Inception module try to explicitly factor two tasks done by a single convolution kernel: mapping cross-channel correlation and spatial correlation
By inception module, these two correlations are sufficiently decoupled due to parallel connections
Would it be reasonable to make a much stronger hypothesis than that of Inception?
Figure 3
We can apply 1x1 convolution over channels and 3x3 convolution separately on each several channels.
That’s somewhat like Depthwise Separable Convolution but in an inverted order!
