My own study says that we cannot top down include "English explanations" of
how the ANNs (Artificial Neural Networks, of which DCNN is just one type)
arrive a conclusions. If you want to translate the computational value of
an ANN into something other than the essential operation that it is
performing, that computational value has to be "grown" into the ANN at the
same time the original computational value is being "reinforced learning
baked in". And when doing that, it costs considerably in both computational
energy and in the extra amount of time the growing of the ANN to produce
the integrated "suggest a move and then offer human meaningful English
explanations for the suggestion". And this assumes English as the language
and a single move suggestion. Add another human language and/or make it
suggest more than one move, and you explode the resources required to
converge on a solution that would eventually beat an amateur, much less a
professional.

Consider ANNs from an entirely different place; our own wetware. Our
wet-ware doesn't learn English and attach explanations to most of its
cognitive activities. And to those activities to which it does attach
English explanations, we have discovered that it is very prone to blind
spots and severe biases that turn into feedback loops magnifying the size
of the blind spots and the degree of biasing. So, even evolution didn't
spend the time to give us a mechanism that can self-describe all or even
most of its operation. And introspection, what we do have that allows us to
self-evaluate is very error prone (apologies to all egos that just got very
activated by being publicly outed as less capable than they know they
actually are).

Another way to consider this is to find out what has happened in the Chess
world with similar desired effects. While they have not been using ANNs
near as strongly, they still have the same desire effects to produce
"English explanations" for move suggestions. I think you will find, even in
this vastly simpler computational space, they haven't made much progress in
this area, either. In otherwords, it is proving to be highly expensive for
insufficient payoff; i.e. an evolutionary dead end.

You say, "Unfortunately no one has a clue on how to put into words what
DCNN "know", to produce really meaningful and useful feedback, justifying
decisions around candidates, etc. This is very much worth investigating."

I have a clue. However, for me personally, I find the investment required
to do said investigation to be WAY too high compared to the actual value
yield I _might_ _eventually_ get from the investment. There is far more low
hanging fruit in the Go AI and ANN space that I could choose (and am
choosing) before I would choose something so highly speculative your
investigation.


On Wed, Mar 30, 2016 at 7:49 AM, djhbrown . <djhbr...@gmail.com> wrote:

> I fully agree with Goncalo that it would be worth investigating how
> one could write an algorithm to express in English what Alpha's or
> DCNNigo's nets
> have learned, and a month ago (before her astonishing achievement in
> March) offerred some ideas on how this might be approached in a
> youtube comment on Kim's review of the Fan Hui games:
> https://www.youtube.com/watch?v=NHRHUHW6HQE
>
> the relevant section of which is (abridged):
>
> "a further, "higher-level" pattern leaning algorithm might be able to
> induce correlation and/or implication relationships between
> convolutions, enabling it to begin to develop its own ontology of
> perceptions, perhaps by correlating convolution relationships with
> geometric patterns on the board image. ... i look forward to the day
> when someone can find a way to induce symbolic pattern descriptions of
> relationships between convolutions and image patterns so that betago
> (child of alpha) can explain its "thinking" in a way we can understand
> and perhaps learn from too."
>
> On 30/03/2016, Gonçalo Mendes Ferreira <go...@sapo.pt> wrote:
> > Come on let's all calm down please. :)
> >
> > David I think the great challenge is in having good insight with AlphaGo
> > strength. Many Faces already provides some textual move suggestions, as
> > do probably other programs. Any program that doesn't use exclusively
> > machine learning or global search, like GNU Go, should be able to
> > suggest how it came about a move.
> >
> > Unfortunately no one has a clue on how to put into words what DCNN
> > "know", to produce really meaningful and useful feedback, justifying
> > decisions around candidates, etc. This is very much worth investigating.
> >
> > - Gonçalo
> >
> >
> >
> > On 30/03/2016 12:32, Álvaro Begué wrote:
> >>> no lack of respect for DeepMind's achievement was contained in my
> >>> posting; on the contrary, i was as surprised as anyone at how well she
> >>> did and it gave me great pause for thought.
> >>>
> >>
> >> Well, you wrote this:
> >>
> >>> but convolutional neural networks and monte-carlo simulators have not
> >>> advanced the science of artificial intelligence one whit further than
> >>> being engineered empirical validations of the 1940s-era theories of
> >>> McCullough & Pitts and Ulam respectively, albeit their conjunction
> >>> being a seminal validation insofar as duffing up human Go players is
> >>> concerned.
> >>>
> >>
> >> That paragraph is disrespectful of AlphaGo and every important
> development
> >> that it was built on. Theorists of the 40s didn't know jackshit about
> how
> >> to make a strong go program or any other part of AI, for that matter.
> >>
> >> This is like giving credit to the pre-Socratic philosophers for atomic
> >> theory, or to Genesis for the Big Bang theory. I am sure there are
> people
> >> that see connections, but no. Just no.
> >>
> >> one has to expect a certain amount of abuse when going public, and to
> >>> expect that eager critics will misrepresent what was said.
> >>>
> >>
> >> Your vast experience in the field means your opinions were formed way
> >> before we knew what works and what doesn't, and are essentially
> worthless.
> >>
> >> There, you like abuse?
> >>
> >> Álvaro.
> >>
> >>
> >> On Wed, Mar 30, 2016 at 6:04 AM, djhbrown . <djhbr...@gmail.com> wrote:
> >>
> >>> one has to expect a certain amount of abuse when going public, and to
> >>> expect that eager critics will misrepresent what was said.
> >>>
> >>> no lack of respect for DeepMind's achievement was contained in my
> >>> posting; on the contrary, i was as surprised as anyone at how well she
> >>> did and it gave me great pause for thought.
> >>>
> >>> as to preconceived notions, my own notions are postconceived, having
> >>> studied artificial intelligence and biological computation over 40
> >>> post-doctoral years during which i have published 50 or so
> >>> peer-reviewed scientific papers, some in respectable journals,
> >>> including New Scientist.
> >>>
> >>> On 30/03/2016, Stefan Kaitschick <skaitsch...@gmail.com> wrote:
> >>>> Your lack of respect for task performance is misguided imo. Your
> >>>> preconceived notions of what intelligence is, will lead you astray.
> >>>>
> >>>
> >>>
> >>> --
> >>> patient: "whenever i open my mouth, i get a shooting pain in my foot"
> >>> doctor: "fire!"
> >>> http://sites.google.com/site/djhbrown2/home
> >>> https://www.youtube.com/user/djhbrown
> >>> _______________________________________________
> >>> 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
> >>
> > _______________________________________________
> > Computer-go mailing list
> > Computer-go@computer-go.org
> > http://computer-go.org/mailman/listinfo/computer-go
>
>
> --
> patient: "whenever i open my mouth, i get a shooting pain in my foot"
> doctor: "fire!"
> http://sites.google.com/site/djhbrown2/home
> https://www.youtube.com/user/djhbrown
> _______________________________________________
> 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

Reply via email to