[Noisebridge-discuss] [ml] I am interested in starting an optimization group at noisebridge (BetterBridge? TrollSearch?)

Sun May 29 02:32:05 UTC 2011

I apologize that perhaps I was all over the place before in this thread. As
far as the algorithms and projects go I am very interested in anything with
longer term really research stuff especially if it might lead to
publication. I'll try my best to get to ML meeting this week if I can.

My interest in ML did indeed evolve from Linuistics/CogSci/Philosopy;
however my formal training was in Math/NumericalScientificComputing. Just I
got away from the latter for some years and coming back in reading a lot got
into the former by self-study. Gradually got more into Computational
linguistics and then out again and exploring some real world problems
interest in the ML group. My more recent math interests mostly in Logic/
reasoning models.

The wiki page for ML at Noisebridge lists a ML video course from Stanford (
see https://www.noisebridge.net/wiki/NBML_Course) , but last week I was
indeed looking at ESL text which Mike included in his refs for random
forests (the first presentation he gave before last week).

My web scraping by hand led me to two or three groups, in the east and
midwest, some connections way back MIT and Yale but more recently Mich,
Indiana and CMU (Europy: a) maybe some Edinburgh school as well ? Used the
term ML back in the late 60's to distinguish them from AI workers b) Luc
Steels' work on evolution of language with Alife and Robotics). Most
striking to me this particular HRI group

http://hrilab.cs.tufts.edu/research/

 and especially impressed by perspective of the CMU Machine Learning dept
leader Tom Mitchell:

Book:
www.cs.cmu.edu/~*tom*/mlbook.html
Interview on his perspective:
videolectures.net/mlas06_*mitchell*_itm/

On Sat, May 28, 2011 at 4:56 PM, Kai Chang <kai.salmon.chang at gmail.com>wrote:

> I did cognitive science at UVA, and there the approach was quite different.
> The curriculum seemed heavily influenced by MIT, with more focus on
> linguistics, pyschology and philosophy. Not sure what the approach at CMU is
> like.
>
> At Stanford, the emphasis seems to be on optimization, classification, etc.
> The problem requires a fitness function or other well-defined metrics to
> test against. More related to math and statistics than cognition in general.
>
> For broader questions of cognition and learning, there are several other
> traditions. Logicians and linguists: Wittgenstein, Godel and Saussure.
> Phenomenologists: Heidegger, Merleau-Ponty, Marx. Philosophers: Hume,
> Nietzsche, Deleuze. Buddhism also has very nuanced views on the subject.
>
> These traditions share doubts about symbolic formalism and rationality as
> sufficient means of describing human intelligence and experience.
>
> For example, we have overwhelming evidence that neurons serve as building
> blocks in networks with the capability to learn broad classes of problems.
> But we still have no foundation to demonstrate that neurons produce our
> phenomenal experience (the rich, inner, subjective world).
>
> Anyways. Crutcher, I'll drop by next week. I got really busy last time ML
> was doing a group project. I'd be interested in hearing about these other
> applications of metaheuristic search too!
>
> On Fri, May 27, 2011 at 8:14 PM, David Faden <dfaden at gmail.com> wrote:
>
>> West = http://www-stat.stanford.edu/~tibs/ElemStatLearn/ ?
>>
>> What's the East Coast book?
>>
>> How would you classify http://aima.cs.berkeley.edu/ ?
>>
>> Interesting to hear about this division. Thanks.
>>
>> On Fri, May 27, 2011 at 7:03 PM, Brian Morris <cymraegish at gmail.com>wrote:
>>
>>> I am very much wanting the more doing / project as-a-group thing.
>>>
>>> I probably won't make it to the ML meeting this coming week tho (most 3/4
>>> weeks I am there)
>>>
>>> There are two perspectives of what Machine Learning is, might call for
>>> lack of better terms the Stanford perspective and the Carnegie-Mellon
>>> perspective (from the location of the authors of two popular texts), or East
>>> and West if you like.
>>>
>>> I generally take the East side and Mike the west,  but also would like to
>>> work on more general problems with a group, maybe  less Statistical Learning
>>> ... and / or problems which have orderings rather than actual or precise
>>> numerical values (ordinal variables), so that optimization is possible even
>>> if formulas cannot be given or numerical data is either too imprecise or
>>> simply unavailable. [Like if you have labels like letter grades, but not
>>> numerically based, how best to assign / design content ?]
>>>
>>> Maybe applies to some problems in knowledge management, natural language
>>> / linguistics problems, policy decision making (how to maximize job
>>> satisfaction for instance), intelligence analysis.
>>>
>>> Kinda short on specific problems, might have to Troll Search them.
>>>
>>> By the way, what's the best way to produce a table of contents rather
>>> than an index (given say a bunch of text scraped off the Web) ?
>>>
>>> On Fri, May 27, 2011 at 1:27 PM, Crutcher Dunnavant <crutcher at gmail.com>wrote:
>>>
>>>> The ML group seems to have grown up quite a bit since the last time I
>>>> paid attention; and I think I should start participating; as the page lists
>>>> many things I'd like to learn and play with.
>>>>
>>>> I am specifically suggesting a group-project oriented group; rather than
>>>> a research group or class. Something that would yield finished projects;
>>>> something where we collaborate on a common code base and problem set, beat
>>>> it to death, publish it (5mof?); and move on to the next one.
>>>>
>>>> There are many applications of metaheuristic search outside machine
>>>> learning; and I don't want to hijack a group which looks healthy.
>>>>
>>>> On Fri, May 27, 2011 at 12:00 PM, Mike Schachter <mike at mindmech.com>wrote:
>>>>
>>>>> Hi Crutcher,
>>>>>
>>>>> I'd be interested in black box optimization. The machine learning
>>>>> group meets up on Wednesdays at 7:30pm in the Church classroom:
>>>>>
>>>>> https://www.noisebridge.net/index.php?title=Machine_Learning
>>>>>
>>>>> Just speaking for myself, I'd be happy to see you share time/space
>>>>> with the ML group to talk about optimization, as it's a core part of
>>>>> machine learning.
>>>>>
>>>>> We don't have anything going on next week, and you're welcome to
>>>>> come in to talk about stuff, I'd be happy to discuss optimization with
>>>>> you!
>>>>>
>>>>>  mike
>>>>>
>>>>>
>>>>>
>>>>> On Fri, May 27, 2011 at 10:10 AM, Crutcher Dunnavant <
>>>>> crutcher at gmail.com> wrote:
>>>>> > I am very interested in starting a black box optimization search
>>>>> group at
>>>>> > noisebridge. This field is called "metaheauristics"; but the name is
>>>>> a
>>>>> > stupid historical artifact (so says everyone in the field).
>>>>> > Optimization is, given a function f(x), searching for the x which
>>>>> yields the
>>>>> > best f(x). Black box optimization is a sub-field of optimization
>>>>> where you
>>>>> > can't analyize the function f to determine what values of x are
>>>>> likely to be
>>>>> > good; so you have to search the space for them.
>>>>> > The following algorithms are ALL metaheuristic optimization:
>>>>> > Hill Climbing (aka. Gradient Assent/Descent)
>>>>> > Genetic Search
>>>>> > Genetic Programming
>>>>> > Ant Colony Systems
>>>>> > Particle Swarm Optimization
>>>>> > I've recently read a fabulous undergraduate text on the subject, very
>>>>> > approachable, called "Essentials of Metaheuristics".
>>>>> > The book in question is available from Lulu and Amazaon:
>>>>> > http://www.cs.gmu.edu/~sean/book/metaheuristics/
>>>>> > or you can just download the PDF.
>>>>> > http://www.cs.gmu.edu/~sean/book/metaheuristics/Essentials.pdf
>>>>> >
>>>>> > If you aren't sure what I'm talking about, read the first chapter or
>>>>> two. If
>>>>> > you have a background in programming, you should be able to follow it
>>>>> > trivially.
>>>>> > What I want TrollSearch to do: Build Shit
>>>>> > Let's find interesting problems; and build search algorithms over
>>>>> them. This
>>>>> > can apply to evolving good fit 3d models for the printer; making
>>>>> techno; or
>>>>> > identifying penii.
>>>>> > I'd like TrollSearch to look much more like SpaceBridge than like the
>>>>> Python
>>>>> > Class.
>>>>> > Please comment in-thread if you are interested.
>>>>> > --
>>>>> > Crutcher Dunnavant <crutcher at gmail.com>
>>>>> >
>>>>> > _______________________________________________
>>>>> > Noisebridge-discuss mailing list
>>>>> > Noisebridge-discuss at lists.noisebridge.net
>>>>> > https://www.noisebridge.net/mailman/listinfo/noisebridge-discuss
>>>>> >
>>>>> >
>>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> Crutcher Dunnavant <crutcher at gmail.com>
>>>>
>>>> _______________________________________________
>>>> Noisebridge-discuss mailing list
>>>> Noisebridge-discuss at lists.noisebridge.net
>>>> https://www.noisebridge.net/mailman/listinfo/noisebridge-discuss
>>>>
>>>>
>>>
>>> _______________________________________________
>>> ml mailing list
>>> ml at lists.noisebridge.net
>>> https://www.noisebridge.net/mailman/listinfo/ml
>>>
>>>
>>
>> _______________________________________________
>> ml mailing list
>> ml at lists.noisebridge.net
>> https://www.noisebridge.net/mailman/listinfo/ml
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.noisebridge.net/pipermail/noisebridge-discuss/attachments/20110528/41c26cba/attachment-0003.html>