When TensorFlow was first released I used it to implement a CNN for move prediction and evaluation, and I requested the addition of a function to implement chain pooling: https://github.com/tensorflow/tensorflow/issues/549
It's now implemented here: https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/unsorted-segment-max By the time they got around to implementing it I wasn't actively doing computer go anymore (I went back to chess for a while), so I haven't actually used it. But it is a very natural idea. Regards, Álvaro. On Fri, Aug 18, 2017 at 2:14 PM, David Wu <lightvec...@gmail.com> wrote: > While browsing the online, I found an interesting idea "chain pooling" > presented here: > https://github.com/jmgilmer/GoCNN > > The idea is to have some early layers that perform a max-pool across > solidly-connected stones. I could also imagine it being useful to perform a > sum. So the input would be a 19x19 layer, and the output would be a 19x19 > layer where the output at a given position, if that position is occupied by > a stone, is equal to the maximum (or the sum of) all the values in the > input layer across all stones that are solidly connected to that group. > > One might imagine going further and allowing the neural net some early > convolutional layers that determine the connectivity strength for this > pooling between groups, so that it could choose to pool across definite > single-point eyes or bamboo joints, etc. It's possible that one would not > want to force all layers through this operation, so possibly only some > feature planes would be fed through this operation, or perhaps all of them > but the identity transformation would also be an output of the layer to > feed into the next. > > Speculatively, in the best case one might imagine this has a chance to > improve the ability of the neural net to evaluate large semeai or to judge > the status of large dragons, by letting it propagate liberty count > information (including virtual liberties due to approach moves) and > information about eyes across the board more rapidly than a series of local > convolutions could do so. In fact, it seems that convolutional layers > followed by an early pooling of this sort would make it unnecessary to > provide liberties as an input feature because it would become easy for the > neural net to compute it on its own, although one would still probably want > to provide it to save the network the effort of having to learn it. > > Of course, this idea could also easily turn out worthless. One thing I'm > very not sure about is how GPU-friendly this kind of operation could be > made to be, since I don't understand GPUs. Any thoughts? > > > > _______________________________________________ > Computer-go mailing list > Computer-go@computer-go.org > http://computer-go.org/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
