Some questions
Graham Higgins
1. Are introductions in order?
2. Noah, when you write """Some of the things I am interested in are
solving "low hanging fruit" problems that can be done this year, or
sooner.""" do you have any examples to share?
3. Is there any interest in creating a Python-oriented code resource
along the lines of the Lisp one maintained [1] by CMU? If so, does
anyone have any preferences / suggestions for a $HOME?
[1] http://www.scs.cmu.edu/afs/cs/project/ai-repository/ai/readme.txt
4. Anyone for SemWeb?
Noah Gift
<gjhi...@googlemail.com> wrote:
>
> Hi all,
>
> 1. Are introductions in order?
the word a bit.
>
> 2. Noah, when you write """Some of the things I am interested in are
> solving "low hanging fruit" problems that can be done this year, or
> sooner.""" do you have any examples to share?
figuring a way out to process more information in less time. The
practical implication is RSS/Atom feeds. There is already enough
information published to get started. The book collective
intelligence has some examples of classifying RSS feeds. Tarek wrote
Atomiser:
http://tarekziade.wordpress.com/2008/12/14/looking-for-beta-testers-for-atomisator/
Phase 1:
In practical terms, I would like to offload the task of reading, say,
Slashdot, Wired, Planet Python, etc, to a bot, that reads it for me,
and then summarizes the content, or classifies the content by
identifying it as something I might want to read.
Phase 2:
Have the bot/agent began to think for me, and proactively find other,
previously unknown, sources of information, and then do the same
thing.
I like this initial problem because it is practical, it could save me
and others 10-60 minutes a day, for example, and it isn't that hard.
My longer term goals is to eventually create an agent that can do
large portions of thinking for me, so I can I think about other
things, but I am prepared that, that problem is a bit more tricky.
>
> 3. Is there any interest in creating a Python-oriented code resource
> along the lines of the Lisp one maintained [1] by CMU? If so, does
> anyone have any preferences / suggestions for a $HOME?
>
> [1] http://www.scs.cmu.edu/afs/cs/project/ai-repository/ai/readme.txt
open source AI solutions in Python in one spot. There is this google
code project from Norvig's book, which might have some good ideas to
look at as well. Maybe if we get traction, we could have a GSOC
project, get Universities involved etc.
http://code.google.com/p/aima-python/
>
> 4. Anyone for SemWeb?
I don't know a lot about the Semantic web, but a friend from Atlanta
is really an expert on it. I have CC'd him, in case he is interested.
Finally, I am very realistic that this might be a boring list, and
some problems might take a few years to solve or more. I don't mind,
I am patient :)
> >
>
--
Cheers,
Noah
Andrew Barilla
Douglas Napoleone
>
> On Fri, Apr 10, 2009 at 6:51 AM, Graham Higgins
> <gjhi...@googlemail.com> wrote:
>>
>> Hi all,
>>
>> 1. Are introductions in order?
>
>
> There might be only me and you so far :) So we might need to spread
> the word a bit.
My name is Doug Napoleone, a research engineer at Nuance
communications working on the MREC large vocabulary continuous speech
recognizer.
>
>>
>> 2. Noah, when you write """Some of the things I am interested in are
>> solving "low hanging fruit" problems that can be done this year, or
>> sooner.""" do you have any examples to share?
>
>
> Yes. The first one I want to solve, a lot for personal reasons, is
> figuring a way out to process more information in less time. The
> practical implication is RSS/Atom feeds. There is already enough
> information published to get started. The book collective
> intelligence has some examples of classifying RSS feeds. Tarek wrote
> Atomiser:
> http://tarekziade.wordpress.com/2008/12/14/looking-for-beta-testers-for-atomisator/
>
> Phase 1:
>
> In practical terms, I would like to offload the task of reading, say,
> Slashdot, Wired, Planet Python, etc, to a bot, that reads it for me,
> and then summarizes the content, or classifies the content by
> identifying it as something I might want to read.
collecting the data into a normalized form will be hard enough.
> Phase 2:
>
> Have the bot/agent began to think for me, and proactively find other,
> previously unknown, sources of information, and then do the same
> thing.
>
> I like this initial problem because it is practical, it could save me
> and others 10-60 minutes a day, for example, and it isn't that hard.
>
>
> My longer term goals is to eventually create an agent that can do
> large portions of thinking for me, so I can I think about other
> things, but I am prepared that, that problem is a bit more tricky.
http://us.pycon.org/2009/conference/schedule/event/71/
replace the jets/sharks data with feed data and you get the idea.
set up a feedback loop for things which are a hit or a miss as a second round.
After that then look at either SLM or FSM based NLU practices for
doing deeper grocking of the article content once you have the initial
feedback (or even a hmm net + verterbi decode). All three require
having a measure of goodness on a subset of known data to properly
train them. The FSM approach might be the best as it directly maps
onto the initial FSM used for initial determination. This would help
in pre-compiling the FSM and reducing it to lower the compute load.
View the NLU FSM as a constraint on the search FSM.
This all assumes a search and nlu fsm per person for a larger system
which could be used for multiple people.
Granted I am viewing this problem through the eyes of speech
recognition, and ultimately everything looks like a decode to me :-)
-Doug
Eric
>> 1. Are introductions in order?
>
>
> There might be only me and you so far :) So we might need to spread
> the word a bit.
Hello,
I am an astronomer in the Nautical Almanac Office at the US Naval
Observatory (though I actually telework from Colorado). I use Python daily
for my work, and--apart from a general interest in AI--I am specifically
interested in algorithms that make working with our large datasets more
efficient, algorithms for finding groups within datasets, and algorithms
that help with optimizing page layouts.
Eric
Ricardo Kirkner
(yet) couldn't find how to make this a full-time occupation (which
would actually be my dream-job).
I have specialized in Reinforcement Learning, during my university
time, and lately I have started to write a pure-python reinforcement
learning library (reply) [0]. I find the proposed problem
of making an agent that is intelligent enough to find out what kind of
stuff is interesting for oneself absolutely intriguing. I would very
much like to see how this comes to life.
Regarding the 'reply' library, you are more than welcome to give it a
look, suggest ideas and improvements, but be warned though, that we
(the reply team) are currently in the middle of a rewrite in order to
make
it compatible with rl-glue [1], so expect things to change in the
following weeks. As said, since this is just a part-time occupation,
progress is not very fast, but happens in 'batches'. I expect we can
deliver a
first release of the library in the following weeks (although most
probably after pyweek [2] has ended).
[0] http://reply.googlecode.com
[1] http://glue.rl-community.org
[2] http://pyweek.org
Carl Trachte
industry; now I am working in software implementation for a defense
contractor (the software has a jython API).
Most of the notes I've seen so far apply to analysis of web data or
language processing. I'm more interested in AI as it applies to
autonomous machines and decision making in specific scenarios. Apart
from a little computational geometry and dabbling a little with Common
LISP, I don't have a lot of experience with AI, or what is
traditionally described as AI.
Thanks.
Carl T.
Mehmet Ali Ertürk
I'm software engineer and also, graduate student from İstanbul, Turkey. I'm using
python in my graduate studies. I'm not a SemWeb expert but working on it.
Thanks,
Mehmet Ali
Graham Higgins
Hash: SHA1
On 9 Apr 2009, at 22:03, Noah Gift wrote:
> So we might need to spread the word a bit.
Given that a rather interesting group is apparently beginning to form,
we might have to think a bit harder about exactly what kind of word we
wish to spread.
I'm an old-school AI-er. My core discipline is psychology, my
specialism is cognitive psychology, my career has been in cognitive
science (AI, model-based programming, knowledge representation & HCI);
initially for a couple of years at Marconi Research and then for a
decade or so at HP Labs Bristol. Since '96 I've been hanging out on
the net as an independent.
On the psych side, I'm mainly interested in mental models; perception,
memory and reasoning - and the cognitive factors in human error. On
the comp side I'm mainly interested in knowledge representation,
reasoning techniques and HCI - and cognitive factors in the design/use
of programming languages.
I'm currently working informally on bringing AI-enhanced decision
support technology within the grasp of the ordinary Joe, the one who
is currently being forced to make woefully under-supported choices
about increasingly more complex matters, the outcome of which will
profoundly affect his future well-being.
>> 2. Noah, when you write """Some of the things I am interested in are
>> solving "low hanging fruit" problems that can be done this year, or
>> sooner.""" do you have any examples to share?
>
>
> Yes. The first one I want to solve, a lot for personal reasons, is
> figuring a way out to process more information in less time. The
> practical implication is RSS/Atom feeds. There is already enough
> information published to get started. The book collective
> intelligence has some examples of classifying RSS feeds. Tarek wrote
> Atomiser:
> http://tarekziade.wordpress.com/2008/12/14/looking-for-beta-testers-for-atomisator/
I agree that it's not too difficult to get started but getting usable
results demands significant effort. Unless you're happy to rely on 3rd
party solutions such as OpenCalais, you're probably going to need to
create your own adequately tagged training corpus.
It is a tempting piece of fruit, though. I decided to limit my goal to
automated tagging and took a very simplistic approach, I'm still
trying to decide whether the result is worth pursuing further [1] (the
output certainly needs tidying). I must admit, RSS items are currently
a lot cleaner and easier to handle than web pages.
Phorm [2, 3 (sections 48-50)] seem to be taking a similar approach (a
frequency count of non stop words) but I can't see that they're going
to be dramatically more successful.
> I like this initial problem because it is practical, it could save me
> and others 10-60 minutes a day, for example, and it isn't that hard.
I think you'll find that the devil is in the detail and in the
training of the classifier.
>> 3. Is there any interest in creating a Python-oriented code resource
> That sounds like a great idea. I like the idea of creating a bunch of
> open source AI solutions in Python in one spot. There is this google
> code project from Norvig's book, which might have some good ideas to
> look at as well. Maybe if we get traction, we could have a GSOC
> project, get Universities involved etc.
>
> http://code.google.com/p/aima-python/
The NLTK [4] is a great resource and I really must make some time to
play with MIT's conceptnet [5].
Noah suggests bitbucket, I'll follow that with knowledgeforge:
http://www.knowledgeforge.net/
>> 4. Anyone for SemWeb?
>
> I don't know a lot about the Semantic web, but a friend from Atlanta
> is really an expert on it. I have CC'd him, in case he is interested.
It might be able to make a contribution to the success of your RSS
summariser. There's some quite interesting stuff knocking around such
as python-dlp [6] which offers DLP-based reasoning over RDF/N3 content
and includes Chimezie Ogbuji's FuXi, a forward-chaining production
system for N3, based on Forgy's Rete algorithm.
I'm pursuing some SemWeb work from a heavily pragmatic perspective - I
want to find out whether it actually works and whether I can use it on
a commercial basis. To that effect, I've chosen a real-world domain,
the UK Parliament, and am attempting to model it using RDF [7]. I'm
looking forward to firing up the inference engines once the domain
model is complete.
> Finally, I am very realistic that this might be a boring list, and
> some problems might take a few years to solve or more. I don't mind,
> I am patient :)
The patience will definitely come in handy but I don't see why this
should necessarily be a boring list, its subscribers thus far seem to
be involved in some interesting work.
Cheers,
Graham
http://www.linkedin.com/in/ghiggins
[1] http://bel-epa.com/resources/ratonit/
[2] http://en.wikipedia.org/wiki/Phorm
[3] http://www.cl.cam.ac.uk/~rnc1/080518-phorm.pdf
[4] http://www.nltk.org/
[5] http://conceptnet.media.mit.edu/
[6] http://code.google.com/p/python-dlp/
[7] http://knowledgeforge.net/semwebparlparse/home.html
-----BEGIN PGP SIGNATURE-----
iEYEARECAAYFAknfMicACgkQOsmLt1NhivxEiwCg33SkoHO64A0FEtjGRNFrx3dl
ugYAoLvMmsxSiPtEXYpPGz45O8zJsFseiQCVAgUBSd8yJ1nrWVZ7aXD1AQISQQP9
E727SAn+KWrEDyKJOsPC18qdIBUViOGUOYrBQ49wK47QJo645eIAOWy7gA5XDKgK
wyiX1Q4n/nTh7jwhsYiEtRkO6M44rMExisFU+YDFzNKHtmS2ZNbkGh7YbI3qv1YI
ZmycWyp68264IL52A+ascao6hZu8iosBOpFy5Rv3cLg=
=GeEZ
-----END PGP SIGNATURE-----
Jeff Rush
> Hi all,
>
> 1. Are introductions in order?
AI. I don't have a strong academic background in the more theoretical
aspects of AI but can provide implementation support, packaging and
distribution energy and organizational input.
And while not precisely AI (by some definition of AI), I also have a
background in the xanalogical storage systems of Ted Nelson,
multidimensional tumblers and time/space trees for organizing mass
quantities of information in a fine-grained way. I'm a cyberarchivist
for the historical records of Project Xanadu and have been
reimplementing his ideas in Python.
> 2. Noah, when you write """Some of the things I am interested in are
> solving "low hanging fruit" problems that can be done this year, or
> sooner.""" do you have any examples to share?
newshound AI. Besides content filtering, I'm also interested in
offloading scheduling and task management, where an AI system can help
me remember to do periodic acts derived from my email/IM streams and
organize task dependencies in a smart way to increase leverage where
certain subtasks advance me on several fronts compared to less
valuable subtasks that I should not pursue.
> 3. Is there any interest in creating a Python-oriented code resource
> along the lines of the Lisp one maintained [1] by CMU? If so, does
> anyone have any preferences / suggestions for a $HOME?
>
> [1]http://www.scs.cmu.edu/afs/cs/project/ai-repository/ai/readme.txt
accumulating a nice library of reusable and documented AI building
blocks. We could obtain the prestigious subdomain of ai.python.org
and I offer my server to host it. I would propose basing the site on
the Sphinx documentation system and a Mercurial version control
system. And I would help to package the modules and get them into the
Python Cheeseshop for easier reuse.
An example of the reference-type presentation that I'm thinking of can
be found at the Sphinx-based site of:
http://www.buildout.org/install.html
http://www.buildout.org/docs/dirstruct.html
http://www.buildout.org/docs/usecase_singlemod.html
> 4. Anyone for SemWeb?
I'm not particular excited by the Semantic Web, seeing it as an
attempt to rewrite the web to make it easier for machines to traverse,
when I think the more pragmatic but difficult approach is to improve
the tools to handle the coarse and disorganized nature of the web we
have.
I'll spread the word of this list to several people I know will be
interested.
-Jeff
Noah
> and I offer my server to host it. I would propose basing the site on
> the Sphinx documentation system and a Mercurial version control
> system. And I would help to package the modules and get them into the
> Python Cheeseshop for easier reuse.
>
Erich
Im Erich. These days I do research work in the critical infrastructure
field (essentially looking for best practices to modernizing power
grid networks and systems). I keep running into these problems whose
solutions t seem like they could benefit from some AI techniques. My
problem is that "seems like" and "do" are very different, but I don't
know enough to apply them. Rewording it: AI looks cool, i wanna
play :)
A couple of general thoughts from stuff I've seen in this thread:
Code: where we put it probably doesn't matter too much, I think if a
"big name" gets picked some people will be happy, some annoyed, for
all cases of "big name". I generally work on the assumption that
anyone who is more concerned with the vcs than the project prolly
shouldn't be on the project anyway :)
Reimplementing the libraries: If these are very high quality and
respected, it may make more sense to provide an interface than a
reimplementation. This of course is ignoring the learning benefits
from reimplementation. However it feels more pythonic to make an
interface and then do really cool things iin python with the
blocks :).
As for Noah's suggested rss agent as a fun project: I'm all in. This
sort of stuff has been promised to me for a very long time. I think
its time to just start taking (and making :) ) it..
Im looking forward to seeing where this group goes.
Regards,
Erich
Rick Thomas
This is the core of an expert system where a large body of if..then..
expressions describe the domain. RETE then provides efficient
inference over the rules.
http://code.google.com/p/python-dlp/
A couple of notes to put this in context:
Think of the W3C semantic web standards RDF and OWL as interchange
formats on which many projects are based. python-dlp allows modeling
in complex domains (the author works in a top cardiology clinic) but
the average semantic wab application is much more prosaic. In fact,
the semantic web is often wrongly criticized for trying to axiomatize
the web along the lines of an expert system. It is better to think of
the semantic web as the foundation of web-scale relational databases.
Given that, then in certain domains fancier logics may apply.
More philosophically, I recommend Andy Clark's book Being There. He
criticizes expert system type AI and related models of human
cognition. Roughly the idea is that human intelligence is not a matter
of logical models in the brain but rather the situated relation of
brain to environment and artifacts - such relations are the
computation so, literally, our minds are not contained in our brains.
By this reasoning our language and our computers are further
extensions of our intelligence, not independent artificial
intelligences. Where Noah says he would like a bot that thinks, I want
a bot to think *with*.
I think it's important to distinguish a least two senses of the term
"artificial intelligence". One is historical describing the research
that lead to expert systems and the type of algorithms the Raymond
Hettinger explains. http://us.pycon.org/2009/conference/schedule/event/71/
By now these are just tools to use when needed.
The other usage is directed to questions of future possibilities,
elusive because they are questions of the nature of the mind.
limno
environmental drivers of infectious disease. We'll be using AI to
decipher global patterns in disease emergence and resurgence. My
interests include: ecology, scientific computing, and global health. I
recently attended Pycon 2009, it was tremendous.
-Nico
Zellyn Hunter
> 1. Are introductions in order?
in checking out neural networks. http://www.cs.toronto.edu/~hinton/,
especially for document classification.
Zellyn
Graham Higgins
Hash: SHA1
On 10 Apr 2009, at 15:34, Rick Thomas wrote:
> the average semantic web application is much more prosaic.
Just in passing ... do you consider that there is such a thing as an
"average semantic web app" at this early stage of adoption?
> More philosophically, I recommend Andy Clark's book Being There. He
> criticizes expert system type AI and related models of human
> cognition. Roughly the idea is that human intelligence is not a matter
> of logical models in the brain but rather the situated relation of
> brain to environment and artifacts - such relations are the
> computation so, literally, our minds are not contained in our brains.
> By this reasoning our language and our computers are further
> extensions of our intelligence, not independent artificial
> intelligences. Where Noah says he would like a bot that thinks, I want
> a bot to think *with*.
>
> I think it's important to distinguish a least two senses of the term
> "artificial intelligence". One is historical describing the research
> that lead to expert systems and the type of algorithms the Raymond
> Hettinger explains. http://us.pycon.org/2009/conference/schedule/event/71/
> By now these are just tools to use when needed.
> The other usage is directed to questions of future possibilities,
> elusive because they are questions of the nature of the mind.
Ach. Given that this group has assembled under an "AI" banner, we
probably ought to ensure that we're in some sort of broad agreement
of the interpretation of the term in this particular Pythonic context.
Taking a pragmatic perspective: I find it hard to believe that
anyone's planning to re-implement Soar or ACT* in Python (but I
suppose it's not beyond the bounds of possibility). Perhaps I'm
mistaken in my perception - does anyone envisage a role for Python in
implementing a model of the kind of situated cognition that Clark is
describing?
I rather saw the primary focus as being on the kind of techniques of
statistical analysis and algorithms referenced by Hettinger and
implemented in Python. It would be these that underpinned Noah's bot.
I suspect that there are quite narrow limits on what we can collate -
the Norvig examples are snugly ensconced on googlecode, similarly Pyke
[1] on sourceforge. My original perception was that there might be a
role for a gradual accretion of more extensive examples and projects
(e.g. an RSS classifier bot) or perhaps, as has been suggested, a
unifying "glue" layer.
<digression>I share many of Clark's views but not his (admittedly
waning) endorsement of the connectionist approach. I recall attending
a talk on the subject given by Geoff Hinton - the neats were
completely sold, the scruffs, along with the psychologists, were
appalled (by and large). At Labs, we took a very close look at the
possibilities but were ultimately dissuaded from adopting it as a
means of classification because of its discontinuity with the
substantial body of empirical evidence from cognitive psychology.
We did hear of a case in the US insurance world where they adopted a
connectionist approach to classifying risk that ran into serious
trouble when it efficiently learned that people with certain zip codes
in their addresses were a greater risk than others and it classified
them accordingly, an illegal form of social exclusion known as "red-
lining". Being a distributed representation, there was little that
could be done either to remedy the problem or prevent the system from
re-learning the relationship.</digression>
Cheers,
Graham
http://www.linkedin.com/in/ghiggins
[1] http://pyke.sourceforge.net
-----BEGIN PGP SIGNATURE-----
iEYEARECAAYFAknfh+sACgkQOsmLt1NhivzpzwCeNG7v6eI6He+tzMQFMO9S1co+
PIUAoPdwKdXNpQZU0mKSLpNqi5wfzIMgiQCVAgUBSd+H61nrWVZ7aXD1AQLWgwP/
YoXiv1LqK2bOmnR7iA1ntvVCZDpazR3nGYJe7W65gJyn1+IvhD75H8k8g6r5Bynp
wNyGdyqURPSWjtw2wh7cLKiJN6T5dF13xakpwdUxTB6TgdSE0jrt9KAg5Jcqv1QU
lTudUG5gfP1noLLg1PRKQoYihv8wbNOk7n01JsAqKxw=
=d0AV
-----END PGP SIGNATURE-----
Petro Verkhogliad
Hi,
my name is Petro Verkhogliad. I am a grad student studying
reinforcement learning (learning automata). I also have an interest in
natural language processing. This group is a great idea.
At one point I actually wrote an RSS reader that classified posts
using Bayesian inference. It would be interesting to work on something
like that again.
Rick Thomas
the average semantic web application is much more prosaic.
Just in passing ... do you consider that there is such a thing as an
"average semantic web app" at this early stage of adoption?
Ach. Given that this group has assembled under an "AI" banner, we
probably ought to ensure that we're in some sort of broad agreement of the interpretation of the term in this particular Pythonic context.
John
> Hi all,
>
> 1. Are introductions in order?
application of AI in engineering decision making - a combination of
expert systems, machine learning, and optimization. I'm especially
interested in the more practical aspects such as letting software
replace reliance on the advice of some individual with "lots of
experience". To me, that's way too unquantifiable and probably not
optimal. I'm actually doing most coding in Ruby these days but I find
both Python and Ruby syntax to be very similar and it's good to share
ideas with smart people whichever language they are using. And I
too, really like some of the examples in "Collective Intelligence".
Erich, while you find yourself involved with the power grid side of
infrastructure, we do more on the civil engineering side (roads, etc).
But I think many of the AI application approaches would be the same.
eswald
> 1. Are introductions in order?
Noah's blog invitation. (Yes, I'm almost a week behind. Yes, I would
like an AI pre-reader.)
> 2. Noah, when you write """Some of the things I am interested in are
> solving "low hanging fruit" problems that can be done this year, or
> sooner.""" do you have any examples to share?
Development Environment [0], whose ostensible goal is to develop
automated players for the Diplomacy board game [1]. I have personally
developed a server and a couple of almost worthless bots, including
one
that attempts to use a horribly huge neural net. Progress on that one
has stalled, given the difficulties in training it properly.
I have taken notice of expert and other problem-solving systems, which
would probably help stalemate lines and other tactical situations, but
I haven't learned enough AI to help with the more generic strategic
decisions, much less interplayer communication.
This is certainly a rich field, but not really low-hanging fruit.
[0] http://www.daide.org.uk/index.xml
[1] http://www.boardgamegeek.com/boardgame/483
> 3. Is there any interest in creating a Python-oriented code resource
> along the lines of the Lisp one maintained [1] by CMU? If so, does
> anyone have any preferences / suggestions for a $HOME?
known to go offline without warning, so I might not be your best
resource. Perhaps if my Parlance server ever makes enough money to be
hosted outside of my basement...
- Eric