On Wed, Oct 18, 2017 at 04:29:47PM -0700, David Doshay wrote:
> I saw my first AlphaGo Zero joke today:
>
> After a few more months of self-play the games might look like this:
>
> AlphaGo Zero Black - move 1
> AlphaGo Zero White - resigns
...which is exactly what my quick attempt to reproduce
I have two questions.
2017 Jan, Master , defeat 60 pros in a row.
2017 May, Master?, defeat Ke Jie 3-0.
Master is Zero method with rollout.
Zero is Zero method without rollout.
Did AlphaGo that played with Ke Jie use rollout?
Is Zero with rollout stronger than Zero without rollout?
Thanks,
This is a quick check of my understanding of the network architecture.
Let's count the number of parameters in the model:
* convolutional block: (17*9+1)*256 + 2*256
[ 17 = number of input channels
9 = size of the 3x3 convolution window
1 = bias (I am not sure this is needed if you are
On 18-10-17 19:50, cazen...@ai.univ-paris8.fr wrote:
>
> https://deepmind.com/blog/
>
> http://www.nature.com/nature/index.html
Another interesting tidbit:
The inputs don't contain a reliable board edge. The "white to move"
plane contains it, but only when white is to move.
So until AG Zero
On Thu, Oct 19, 2017 at 11:04 AM, Hiroshi Yamashita
wrote:
> I have two questions.
>
> 2017 Jan, Master , defeat 60 pros in a row.
> 2017 May, Master?, defeat Ke Jie 3-0.
>
> Master is Zero method with rollout.
> Zero is Zero method without rollout.
>
> Did AlphaGo that
Well, if you have both, why not use both :)
On Thu, Oct 19, 2017 at 11:51 AM Richard Lorentz
wrote:
> An interesting juxtaposition.
>
> Silver said "algorithms matter much more than ... computing".
>
> Hassabis estimated they used US$25 million of hardware.
>
An interesting juxtaposition.
Silver said "algorithms matter much more than ... computing".
Hassabis estimated they used US$25 million of hardware.
___
Computer-go mailing list
Computer-go@computer-go.org
What shall I say?
Really impressive.
My congratulations to the DeepMind team!
> https://deepmind.com/blog/
> http://www.nature.com/nature/index.html
* Would the same approach also work for integral komi values
(with the possibility of draws)? If so, what would the likely
correct komi for 19x19
Yes, it seems really odd that they didn't add a plane of all ones. The
"heads" have weights that depend on the location of the board, but all the
other layers can't tell the difference between a lonely stone at (1,1) and
one at (3,3).
In my own experiments (trying to predict human moves) I found
The order of magnitude matches my parameter numbers. (My attempt to
reproduce a simplified version of this is currently evolving at
https://github.com/pasky/michi/tree/nnet but the code is a mess right
now.)
On Thu, Oct 19, 2017 at 07:23:31AM -0400, Álvaro Begué wrote:
> This is a quick check
I would like to know how much handicap the Master version needs against the
Zero version. It could be less than black without komi or more than 3 stones.
Handicap differences cannot be deduced from regular Elo rating differences,
because it varies depending on skill (a handicap stone is more
Sure, both hardware and software / algorithms are needed... but which gets
you the bigger ROI? { Just a rhetorical question, I know it is not linear
and not a simple question... but in general, I can see David Silver's (&
Richard Lorentz / Demis Hassabis' counter) point }.
May you live in sente,
Cost reduction in IC has reached or is reaching its limits. Intels 5n techk
is not really a 5n and 5n is not really reachable. Not at least without
some seriously new physics and even then there will be hard limits like
quantum un--certainty. This particular chip may get cheaper if it is ever
done
So I am reading that residual networks are simply better than normal
convolutional networks. There is a detailed write-up here:
https://blog.waya.ai/deep-residual-learning-9610bb62c355
Summary: the residual network has a fixed connection that adds (with no
scaling) the output of the previous
Yes, residual networks are awesome! I learned about them at ICML 2016 (
http://kaiminghe.com/icml16tutorial/index.html). Kaiming He's exposition
was fantastically clear. I used them in my own attempts at training neural
networks for move prediction. It's fairly easy to train something with 20
On 19.10.2017 20:13, Richard Lorentz wrote:
Silver said "algorithms matter much more than ... computing".
Hassabis estimated they used US$25 million of hardware.
Today, it seems 4 TPU cost US$25 million. In 5 or 10 years, every
computer might have its 4-TPU-chip costing $250, if not $25. At
So there is a superstrong neural net.
1) Where is the semantic translation of the neural net to human theory
knowledge?
2) Where is the analysis of the neural net's errors in decision-making?
3) Where is the world-wide discussion preventing a combination of AI and
(nano-)robots, which
17 matches
Mail list logo