Inception v2 and v3 were released together and propose several improvements over the original Inception block (Rethinking the Inception Architecture for Computer Vision, https://arxiv.org/abs/1512.00567). The first is the factorization of the 5×5 convolution in two stacked 3×3 convolutions. We discussed the advantages of this in the Introduction to Visual Geometry Group section.
We can see the new Inception block in the following diagram:
The next improvement is the factorization of an n×n convolution in two stacked asymmetrical 1×n and n×1 convolutions. For ...