How to build a non-full-connection networks


I’m trying build a networks like the pic show.You see the inputs don’t connect the 1st hidden layer with full connection.I have no idea how to make it.Could some one please do me a favor?

If you want the same weights, you could use strided convolution. If not, you could use unfold/fold to get it into a format where multiplying with the same weights would be convolution and do a batch matrix multiplication with different weights.

Quite likely it is useful to make more precise (equations or so) how your input looks and how you want to apply connection weights to it.

Best regards

Thomas

Thank you for your notion.Convolution is a good choise.