Overlapping blocks

[flare9x]

In convolution neural networks the block moves in step sizes (overlaps). It would be a useful feature to have the option to split a matrix by overlapping block sizes:

http://cs231n.github.io/convolutional-networks/

[dlfivefifty]

Could you explain a bit more?

[flare9x]

http://deeplearning.stanford.edu/wiki/index.php/File:Convolution_schematic.gif

See the left image yellow box moves in 1 step increments - it moves horizontal at 1 step increments and once passed to end of the image (array) steps down 1 row and begins sliding through the image to the end.

One can specify step sizes... this image is 1 step increments.

So was thinking to have a feature like that - where the matrix could be populated that way - it might become very big though. Not sure ow feasible but wanted to think about it here.

[dlfivefifty]

I would need to see this in actual mathematics to understand what you want and how BlockArrays.jl fits in.

[flare9x]

https://discourse.julialang.org/t/what-is-julias-im2col/14066

This type of arrangement solves the Q.