Summary on Shusen Wang's video of: DCN (Deep & Cross Network)

 DCN (Deep & Cross Network)  is used to replace fully connected layer.

Cross Layer:

- Taking an intermediate output vector xi, and pass it through a fully connected layer to receive y of the same size as input x0. Take a Hadamard product of x0 and y, and call it z. Let Xi+1 = z + xi be the output.

Xi+1 = x0  * [W * xi + b] + xi

Cross Network:

- Apply cross layer multiple times with the same x0

Deep & Cross Network:

- Take all features as a vector to feed in a fully connected layer and a second path of Cross Network. Concatenate the two results and feed into another fully connected layer.

- widely used in industry for recommendation - to sort the items, or used in shared bottom.

Reference: https://www.youtube.com/watch?v=yNeRO5m63JQ