Am 10/21/2017 um 03:12 AM schrieb uurtamo .:
This sounds like a nice idea that is a misguided project.
[...]
Just accept that something awesome happened and that studying those
things that make it work well are more interesting than translating
coefficients into a bad understanding for people.
I'm sorry that this NN can't teach anyone how to be a better player
through anything other than kicking their ass, but it wasn't built for
that.
Roberts approach might be misguided, but I don't agree that having the
raw network data couldn't teach us something. E.g. have a look at this
guy who was able to identify the neurons responsible for generating URLs
in a wikipedia text generating RNN:
http://karpathy.github.io/2015/05/21/rnn-effectiveness/#visualizing-the-predictions-and-the-neuron-firings-in-the-rnn.
E.g. it might be possible to find the network Part of AlphaGo Zero which
is responsible for L&D problems and use it to dream up new Problems! The
possibilities could be endless. This kind of research might have been
easier with the "classic" AlphaGo with separated policy and value
networks, but should be possible anyways.
Also lets not forget DeepMinds own substantial research in this area:
https://deepmind.com/blog/cognitive-psychology/.
I understand that DeepMind might be unable to release the source code of
AlphaGo due to policy or licensing reasons, but it would be great (and
probably much more valuable) if they could release the fully trained
network. As Gian-Carlo Pascutto has pointed out, replicating this would
not only incur high hardware costs but also take a long time.
David O.
On Fri, Oct 20, 2017 at 8:24 AM, Robert Jasiek <jas...@snafu.de
<mailto:jas...@snafu.de>> wrote:
On 20.10.2017 15:07, adrian.b.rob...@gmail.com
<mailto:adrian.b.rob...@gmail.com> wrote:
1) Where is the semantic translation of the neural net to
human theory
knowledge?
As far as (1), if we could do it, it would mean we could
relate the
structures embedded in the net's weight patterns to some other
domain --
The other domain can be "human go theory". It has various forms,
from informal via textbook to mathematically proven. Sure, it is
also incomplete but it can cope with additions.
The neural net's weights and whatnot are given. This raw data can
be deciphered in principle. By humans, algorithms or a combination.
You do not know where to start? Why, that is easy: test! Modify
ONE weight and study its effect on ONE aspect of human go theory,
such as the occurrance (frequency) of independent life. No effect?
Increase the modification, test a different weight, test a subset
of adjacent weights etc. It has been possible to study semantics
of parts of DNA, e.g., from differences related to illnesses.
Modifications on the weights is like creating causes for illnesses
(or improved health).
There is no "we cannot do it", but maybe there is too much
required effort for it to be financially worthwhile for the "too
specialised" case of Go? As I say, a mathematical proof of a
complete solution of Go will occur before AI playing perfectly;)
So far neural
nets have been trained and applied within single domains, and any
"generalization" means within that domain.
Yes.
--
robert jasiek
_______________________________________________
Computer-go mailing list
Computer-go@computer-go.org <mailto:Computer-go@computer-go.org>
http://computer-go.org/mailman/listinfo/computer-go
<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
_______________________________________________
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
