Any use of non-rectangular-shaped kernels in convolutional neural networks? Especially when analyzing game boards I have been reading a pile of papers on convolutional networks and reinforcement learning.  
I remember seeing an important paper with a non-rectangular shape of the convolution layer (the green shape in this silly drawing). But now I cannot find it.

It might have been something similar to the AlphaGo paper or reinforced learning on a game board.
Can any one suggest or guess which paper it was?
 A: {1} compared square versus triangular 2D convolutions

As Geomatt22 mentions, in the example you gave the question, one could use a square filter and hope that the "actual" shape of the filter be  learnt during the training phase.

{1} Graham, Ben. "Sparse 3D convolutional neural networks." arXiv preprint arXiv:1505.02890 (2015). https://scholar.google.com/scholar?cluster=10336237130292873407&hl=en&as_sdt=0,22 ; https://arxiv.org/abs/1505.02890
A: This seems to come up in earlier Herbrich papers on Go.


*

*"Learning on Graphs in the Game of Go" - where he looks at the board as a different topology

*And this slide in a 2015 presentation he makes, mentioning 13 different "patterns"
(which is somewhat different that the AlphaGo approach)

References


*

*Graepel, T., Goutrie, M., Krüger, M., & Herbrich, R. (2001, August). "Learning on graphs in the game of Go." In International Conference on Artificial Neural Networks (pp. 347-352). Springer Berlin Heidelberg.

*Herbrich, R. (2015) "Machine Learning in Industry". Retrieved from http://mlss.tuebingen.mpg.de/2015/slides/herbrich/herbrich.pdf
