There are few conference available with Jerome Pesenti, Vice President
of Watson Core Technologie, who talks about techniques inside Watson.
Jerome Pesenti said at a conference (at Paris Tech, a french engineering
school, date unknown ~09/2015) that :
- Watson did not use deep learning in the jeopardy version
- But the system evolves continously, they are replacing many things in
Watson by deep learning.
He said that they are replacing codes in the jeopardy version by deep
learning because it's much more efficient in natural language processing
and others. With deep learning, there will soon be a version of jeopardy
for other languages than English.
LAU
Le 10/01/2016 15:02, Jim Bromer a écrit :
I am interested in the question of whether Watson used deep learning
in the Jeopardy version because I am skeptical that there is a clear
cut distinction between hybrids of computational methods that train on
large corpora of data and Deep Learning. A few lines in a computer
review does not convince me. Are you saying (for instance) that the
statistical analysis that was used in Watson was not "Deep"? How could
you know? What are the differences?
There are times when editorial criticisms are useful and there are
times when they are trivial to the issue being discussed. If I asked
people to read something that I posted on my website or which read
like I might try to get it published then I probably would appreciate
comments about typos and grammatical issues. Some time ago someone
pointed out that I was using the word 'discreet' when the word should
have been 'discrete'. I appreciated knowing that I was making that
mistake because it is important to the subject being discussed and I
kept repeating the mistake. However, he also made some put down
suggesting that the fact that I was not using the word 'discrete' when
I meant 'discrete' showed that I did not know too much about computer
programming. I disagree with that point of view because the word
'discreet' is, in my opinion, a very important concept in psychology
and a major problem in contemporary AI. I think contemporary AI
programs lack discretion when confronted by interpretations that might
take multiple paths. So while my mistake was a serious one (when
talking in a computer group) it was not an indication that I did not
have too much experience thinking about the subject of this group. A
lack of discretion can be taken as a lack of insight,
but the psychology of discretion is, in my opinion, something that is
very seriously lacking in contemporary AI. Narrow AI can show some
discretion as long as the problem is within the narrow range and the
response is within the range of appropriate responses.
I would be interested in following up on the question of how
Jeopardy's Watson, which the reviewer said uses statistical analysis
on vast unstructured piles of text is essentially different from Deep
Learning.
The chess playing programs are narrow but similar methodologies can be
used for situations where 'positions' can be evaluated so the
underlying methods have much broader general applications.
Conceptual integration is a thing that is very important to me.
However, I do not have it all figured out.
But I can look at (simple) computational analyses and see, for
example, that not all the parts in an algorithm are alike. Operations
can be numbered and they can even - to some extent - be used in
numerical processes. However, that does not mean they are the same or
can then be used in the same way as the numerical operands of the
function. So here you might see that knowledge about an operand and
an operation In a computer function can be useful as long as that
knowledge is then integrated in a suitable way. For another simple
example, people will sometimes try to take the enumeration of a column
of the digits in an arithmetic problem and treat is as if it were a
digit (of one of the operands). (A n-ary number will consist of digits
in columns. For binary the columns are the ones column, the twos
column, the fours column and so on.) Using the ordinal value of a
column might workout in some cases but in others it might not because
the ones in the columns will stand for 1 or 2 or 4 or 8 and so on. So
you have to keep track that an explicit enumeration of the columns may
have more than one meaning in an algorithm. Suppose that this was the
first time someone ever considered this problem. In order to make
sense of this he would have to be able to integrate a number of very
simple concepts. Even if someone is capable of understanding the
simple concepts when they are taken out of context (what do I mean by
a column of a number, what do I mean by a digit, what do I mean by an
n-ary number, what do I mean by the enumeration of the columns of a
number can take on different meanings) they still might be totally
baffled by what I am talking about. Not only do they have to integrate
these different simple concepts they also have to do so in a very
discreet way. They would need to try to integrate the concepts in
different ways but show great discretion in limiting the number of the
ways that they tried. Just mashing all the concepts together and
trying to make them all act like they were of the same kind of thing
(a countable digit in this example) isn't going to cut it.
Jim Bromer
On Sat, Jan 9, 2016 at 11:00 PM, John Smith <[email protected]
<mailto:[email protected]>> wrote:
"The idea that Deep Blue and Watson were not cases of Deep
Learning is irrelevant. (You are effectively criticizing my topic
headline rather than what I was getting at.)"
Maybe you shouldn't have a title that says one thing while
intending something else?
"But, Deep Learning is being used in visual recognition and my
feeling is that since Watson did use machine learning I believe
that it must have used something that had some correspondence to
Deep Learning."
Your feeling is wrong, Watson didn't have deep learning when it
won jeopardy, it was only added recently
http://www.technologyreview.com/news/539226/ibm-pushes-deep-learning-with-a-watson-upgrade/
There are many kinds of machine learning that are different in
kind from deep learning.
"The argument that they were just narrow AI is also irrelevant."
No it isn't, because narrow AI.. like a machine specifically
designed to play chess, will not be able to do something like play
checkers, or drive a car, or write poetry. It will only be able
to play chess.
"There is no question that Watson and methodologies that are on
par with contemporary Deep Learning have a wide variety of
applications." You know duck tape has lots of applications too..
"So they are capable of some generalization."
Again a chess playing machine can't do jack, but play chess, so
too with a jeopardy playing machine.
"Human beings, which represent the model of general intelligence,
are not capable of figuring out many kinds of problems including
many that computers can and will solve. "
This is the first true thing you've said.
"The problem is that these contemporary AI programs are not
capable of integrated general intelligence and they are end up
working within relatively narrow fields."
Okay first of all please use proper grammar. Second of all what
is this "integrated" general intelligence you speak of, please
define, and please keep in mind I'm a very simple person who has
difficulty with obscure terminology that is only understood in the
mind of the speaker.
"But to say that they are narrow as opposed to genera is not quite
right."
So if someone creates an AI for playing chess and only chess.. it
isn't narrow because you believe there are other applications for
it? This is just wrong. The only way it would have other
applications is if you spent the time to some how map your other
application onto a chess board. But that isn't the AI doing the
generalizing, rather it is the user doing the generalizing.
Narrow AI != General AI
QED
On Sat, Jan 9, 2016 at 10:19 PM, Jim Bromer <[email protected]>
wrote:
The idea that Deep Blue and Watson were not cases of Deep
Learning is irrelevant. (You are effectively criticizing my
topic headline rather than what I was getting at.) But, Deep
Learning is being used in visual recognition and my feeling is
that since Watson did use machine learning I believe that it
must have used something that had some correspondence to Deep
Learning.
The argument that they were just narrow AI is also irrelevant.
There is no question that Watson and methodologies that are on
par with contemporary Deep Learning have a wide variety of
applications. So they are capable of some generalization.
Human beings, which represent the model of general
intelligence, are not capable of figuring out many kinds of
problems including many that computers can and will solve. The
problem is that these contemporary AI programs are not capable
of integrated general intelligence and they are end up working
within relatively narrow fields. But to say that they are
narrow as opposed to genera is not quite right.
Jim Bromer
On Sat, Jan 9, 2016 at 8:31 PM, John Smith
<[email protected]> wrote:
"winning at chess (IBM Deep Blue [doesn't use deep
learning]), recognizing objects in pictures (Many
Companies and different algorithms [some just use
mechanical turk]) and winning at jeopardy (IBM Watson
[didn't use deep learning when it won at jeopardy])."
So none of those achievements used deep learning.
Google's deep mind hasn't "solved intelligence" yet, so it
would be a mistake to expect the kinds of advanced search
capabilities you are thinking of.
IBM did the Jeopardy grand challenge specifically because
they saw Ken Jennings winning streak and the amount of
attention it was attracting, and they thought if we create
a software system that could do that we would get a great
deal of attention, which I'm sure they thought would
subsequently lead to big contracts. So yes it was in a
way a publicity stunt from its inception. And since the
algorithms were hand crafted for a single end (win at
Jeopardy) of course it wasn't going to have a large impact
on the field of AGI in general! Watson wasn't AGI, it was
the waste of time/money narrow AI that the short sighted
people in industry find easy to sell.
On Sat, Jan 9, 2016 at 3:34 PM, Jim Bromer
<[email protected]> wrote:
The hype and the implied conquest of AI that winning
at chess,
recognizing objects in pictures and winning at
jeopardy seems to imply
just does not jive with the fact that search engine
technology lacks
any noticeable intellect even though the computing
power that Google,
Bing or IBM and thousands of other corporations
possess is extremely
impressive.
Jim Bromer
On Sat, Jan 9, 2016 at 3:29 PM, Jim Bromer
<[email protected]> wrote:
> If industry has AI pretty well figured out then why
are search engines
> so incapable of thinking outside the box? The
conclusion looks
> inescapable to me. Yes there will be a day when
someone makes a
> significant achievement while the rest of us might
miss it completely
> but the idea that contemporary deep search (or some
other AI method)
> has achieved the hype or the implied conquest that
winning at chess
> and jeopardy seems to imply just does not jive with
the computing
> power Google, Bing or IBM have. There is a
substantial disconnect
> between low level -almost- human reasoning and deep
learning.
> Jim Bromer
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed:
https://www.listbox.com/member/archive/rss/303/26973278-698fd9ee
Modify Your Subscription:
https://www.listbox.com/member/?&;
Powered by Listbox: http://www.listbox.com
*AGI* | Archives
<https://www.listbox.com/member/archive/303/=now>
<https://www.listbox.com/member/archive/rss/303/24379807-653794b5>
| Modify <https://www.listbox.com/member/?&;> Your
Subscription [Powered by Listbox] <http://www.listbox.com>
*AGI* | Archives
<https://www.listbox.com/member/archive/303/=now>
<https://www.listbox.com/member/archive/rss/303/26973278-698fd9ee>
| Modify <https://www.listbox.com/member/?&;> Your
Subscription [Powered by Listbox] <http://www.listbox.com>
*AGI* | Archives <https://www.listbox.com/member/archive/303/=now>
<https://www.listbox.com/member/archive/rss/303/24379807-653794b5>
| Modify <https://www.listbox.com/member/?&;> Your Subscription
[Powered by Listbox] <http://www.listbox.com>
*AGI* | Archives <https://www.listbox.com/member/archive/303/=now>
<https://www.listbox.com/member/archive/rss/303/27172223-36de8e6c> |
Modify
<https://www.listbox.com/member/?&;>
Your Subscription [Powered by Listbox] <http://www.listbox.com>
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
