The alphaGo network is detailed in their paper.  They have about 50 binary 
inputs, one layer of 5x5 convolutional filters, and about 12 layers of 3x3 
convolutional filters.  Detlef’s net is specified in the prototxt file he 
published here.  It’s wider and deeper, but with fewer inputs.

 

The current popular approach is a to use 5x5 or 7x7 filters in the first 
layers, and 3x3 filters in higher layers.  The topmost layer is special, 
typically a 1x1 convolutional filter.  AlphaGo uses position dependent biases. 
RelU seems to work well, and pooling is not used (for obvious reasons).

 

In my experiments it is essential to have the first layer filters larger than 
3x3.  In higher layers 3x3 seems to work fine.

 

Hope this helps.

 

David

 

From: Computer-go [mailto:[email protected]] On Behalf Of 
Urban Hafner
Sent: Friday, May 20, 2016 4:47 AM
To: [email protected]
Subject: [Computer-go] Commonly used neural network architectures

 

Hey there,

 

just like everyone else I’m currently looking into neural networks for my go 
program. ;) Apart from the AlphaGo paper where I can I find information about 
network architecture? There’s the network from April 2015 from Detlef 
(http://computer-go.org/pipermail/computer-go/2015-April/007573.html) but I 
don’t know enough about caffe to figure out the architecture. Basically, I more 
interested in understanding how to build a network myself than just using a 
pre-trained network.

 

Cheers,

 

Urban

-- 

Blog: http://bettong.net/

Twitter: https://twitter.com/ujh

Homepage: http://www.urbanhafner.com/

_______________________________________________
Computer-go mailing list
[email protected]
http://computer-go.org/mailman/listinfo/computer-go

Reply via email to