Page 3 of 3
Re: LZ v Golaxy
Posted: Thu Apr 26, 2018 11:35 am
by abcd_z
Hane wrote:Golaxy， while adopting AlphaGo‘s basic architecture， [...] manages to complete model training with less computing resources or training samples
You know what that says to me? He used transfer learning from Leela Zero's weights.
Re: LZ v Golaxy
Posted: Thu Apr 26, 2018 4:36 pm
by dfan
abcd_z wrote:
You know what that says to me? He used transfer learning from Leela Zero's weights.
I don't know about the weights, but I think at the very least it would be silly not to use Leela Zero's millions of quality self-play games when training any network.