Post-doc in Acoustic Event Detection (US citizens only)

tony.nospam@nospam.tonyRobinson.com wrote:
> Rune,
>
> I wrote a long reply which on re-reading of it has turned into somewhat
> of a rant, so I'm not going to post it and see instead if we can agree
> on what's causing you so much upset.
>
> I think you are saying that you view the proposed task as a comparison
> over all possible signals where you view the pattern space as vast and
> impossible to enumerate and so you feel there will always be some part
> of the pattern space that is poorly modelled.

Agreed.

> I view the task as one of
> a comparison over pattern space for which it is possible to come up with
> realistic model of the distribution.  The space may still be vast and
> impossible to enumerate but over the distribution considered sensible
> models may be formulated and evaluated.
> Do we agree on this?

I agree that this is a sensible way of attacking the problem. Our
disagreement is on whether it is this approach that is outlined in
the ad, and whether it is this approach the applicant will be
assigned to work with.

If you are certain your description represents the work at your
institution, I suggest you should have a quiet word with whoever
wrote that ad and published it as an official ICSI document.

> Otherwise you finished your last post with an
> indefensible libellous on ICSI, so I shall don my flameproof suit...

Maybe indefensible what ICSI is concerned, as I have never worked
there, but completely defensible with respect to the majority of my
previous projects and employers. There ARE people around who
regard "compare the measured data to all possible models,
each with all possible parametrizations" as a sane, viable way to
analyze data. There ARE people -- distingushed professors with
dozens of publicatuions and decades of tenure, no less -- who regard
detonating on the order of 50-100 kg of TNT on a regular basis as
a perfectly viable (if somewhat expensivee as one would have to
use P3 aircraft to deploy the charges) preparation for PASSIVE
sonar.

I've worked with both approaches. No fun. No fun whatsoever.

Rune

"Rune Allnor" <allnor@tele.ntnu.no> writes:

> tony.nospam@nospam.tonyRobinson.com wrote:
> 
> > I view the task as one of
> > a comparison over pattern space for which it is possible to come up with
> > realistic model of the distribution.  The space may still be vast and
> > impossible to enumerate but over the distribution considered sensible
> > models may be formulated and evaluated.
> > Do we agree on this?
> 
> I agree that this is a sensible way of attacking the problem. Our
> disagreement is on whether it is this approach that is outlined in
> the ad, and whether it is this approach the applicant will be
> assigned to work with.

Here is the original wording again: "The Fellow will be working with
Nelson Morgan, along with international colleagues, in the area of the
detection of one particular target word and/or sound in the background
of all other possible words or any other realistic sounds."

The "all other possible" refers to words, and we are agreed that it is
possible to make reasonable approximations to the distributions of word
acoustic features (that's how large vocabulary speech recognition
works).  I think the use of the word "realistic" in the second part,
"any other realistic sounds", implies that it's also possible to make
reasonable approximations to the distributions of the non-word acoustic
features.  I don't think that this needs to be made any more explicit in
a job advert.

However, you've obviously had some bad experiences which has fired you
up on this topic.  On the other hand I've had a very good time working
with Morgan and the ICSI team.  So naturally there is plenty of fuel on
both sides for full scale ranting here.  I'm very confident that your
fears are unfounded in this case, but after what you've said then, yes,
there are probably research projects running out there with unrealistic
goals, and yes, that is no fun whatsoever.



Tony

Wasting taxpayer money ?

OK, keep wasting it for now... until someone comes up with the perfect
feature extractor (a.k.a "artificial ear") for speech recognition. Then
you are out of work...

dave_gelbart@yahoo.com wrote:
> The International Computer Science Institute (ICSI) invites
> applications for a postdoctoral Fellow position in acoustic event
> detection. The Fellow will be working with Nelson Morgan, along with
> international colleagues, in the area of the detection of one
> particular target word and/or sound in the background of all other
> possible words or any other realistic sounds. The applicant is expected
> to have strong capabilities in signal processing. Familiarity with
> speech algorithms per se is a plus but is not absolutely required.
>
> ICSI is an independent not-for-profit institute located a few blocks
> from the University of California Berkeley campus. It is closely
> affiliated with the University, and particularly with the EECS
> Department. It has world-class activities in Internet research, natural
> language processing, bioinformatics, and speech processing. See
> www.icsi.berkeley.edu to learn more about ICSI.
>
> The ICSI Speech Group (including its predecessor, the ICSI Realization
> Group) has been a source of novel approaches to speech processing since
> 1988. It is primarily known for its work in speech recognition,
> although it has housed major projects in speaker recognition and
> metadata extraction in the last few years. The new effort will draw
> upon lessons learned in our feature extraction work for speech
> recognition.
>
> Applications should include a cover letter, vita, and the names of at
> least 3 references (with both postal and email addresses). Applications
> should be sent by email to morgan@icsi.berkeley.edu or by postal mail
> to:
>
> Director (re Acoustic Event Detection Postdoctoral Search)
> ICSI
> 1947 Center Street Suite 600
> Berkeley, CA 94704
>
> ICSI is an Affirmative Action/Equal Opportunity Employer. Applications
> from women and minorities are especially encouraged.
> 
> For this particular position, US citizenship is required.

"Rune Allnor" <allnor@tele.ntnu.no> wrote in news:1156489566.536459.51370@
75g2000cwc.googlegroups.com:

> The term "all possible" is insane from both an engineering and a R&D
> point of view. How do you know that you have covered **all**possible**
> variants of anything?



"All possible" of course approaches hyperbole, but its a rather convenient 
term to make your goals understood in a short sentence in lay language, 
which is what's usually required if you expect to successfully attract 
funding from government sources.

-- 
Scott
Reverse name to reply

Rune Allnor wrote:

   ...

> The term "all possible" is insane from both an engineering and a R&D
> point of view. How do you know that you have covered **all**possible**
> variants of anything? How do you know that some person from half-way
> around the world doesn't throw in the odd word of his native language?
> There are some languages of the bushmen in the Namib desert that
> sound like clicks to me and most probably to almost everybody else
> in the world. Do you count these "clicks" among the "all possible
> words" of yours? If not, on what basis do you discriminate between
> "possible" and "impossible" words?
> 
> Paying attention to detail is the key characteristic of the
> professional,
> irrespective of trade or craft.

I'm reminded of a line from the movie /Cool Hand Luke/: "What we have 
here is a failure to communicate." These PR types get so caught up in 
(what they think is) fancy rhetoric that they lose sight of the actual 
meaning of what they write. It's such a widespread failing that most of 
the time it slides by without anyone noticing. I applaud Rune for 
picking up on it this time. I'm tired of guessing at probable meanings.

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

tony.nospam@nospam.tonyRobinson.com wrote:
> "Rune Allnor" <allnor@tele.ntnu.no> writes:
>
> > tony.nospam@nospam.tonyRobinson.com wrote:
> >
> > > I view the task as one of
> > > a comparison over pattern space for which it is possible to come up with
> > > realistic model of the distribution.  The space may still be vast and
> > > impossible to enumerate but over the distribution considered sensible
> > > models may be formulated and evaluated.
> > > Do we agree on this?
> >
> > I agree that this is a sensible way of attacking the problem. Our
> > disagreement is on whether it is this approach that is outlined in
> > the ad, and whether it is this approach the applicant will be
> > assigned to work with.
>
> Here is the original wording again: "The Fellow will be working with
> Nelson Morgan, along with international colleagues, in the area of the
> detection of one particular target word and/or sound in the background
> of all other possible words or any other realistic sounds."
>
> The "all other possible" refers to words, and we are agreed that it is
> possible to make reasonable approximations to the distributions of word
> acoustic features (that's how large vocabulary speech recognition
> works).

These are three statements in one sentence:

1) The scope of ambition ("all possible words")
2) The reasonable characterization of sound signals
3) The characterization of "word acoustic features"

As far as I recall, I have agreed to only one of these premises,
item 2), that it is reasonable to define features of acoustic signals
that may turn out to be useful in some sort of analysis.

As for item 1), it ought to be clear by now that I regard
that stype of scope as insane.

I have never agreed to item 3, which I can only interpret to
state that there is some characteristic unique to human
words, as opposed to both non-anthropogenic sounds or
anthropogenic sounds that are not words. Please refer to
any statement of mine you might have interpreted as such
an agreement, and we will sort out the misunderstanding.

As for vocabulary speech recognition, please refer to any
system that does *not* take advantage of a severely
constrained data recording. I only know of vocabulary
speech recognition systems where the signal to be
processed is *known* to be of a human voice, like in
a telephone or where the user wears a microphone.

The restriction of the data recording to contain almost
solely human speech, is one of those tiny details which
implications a theoretician with no relevant practical
experience would understand.

> I think the use of the word "realistic" in the second part,
> "any other realistic sounds", implies that it's also possible to make
> reasonable approximations to the distributions of the non-word acoustic
> features.  I don't think that this needs to be made any more explicit in
> a job advert.

The job advert is as explicit as it can get in that the issuing
institution is not by any means up to scratch, professionally
speaking:

- Somebody drafted an ad that not by any stretch of the
  imagination can be said to be realistic in scope
- Somebody did not catch this glitch during proofreading
- The ad needs considerable interpretation, goodwill and
  sheer incompetence with the reader to be regarded as
  interesting
- The ad projects an image to the general public about what
  these people work with, that isn't by any means realistic
- The ad confirms various views held by laymen about what
  it is reasonable to expect from speech processing technology
  in general, and this institution in particular

Every time I have got myself into a really bad project, my instincts
and alarm bells have trigged on far more subtle clues than these. 

Rune

"Rune Allnor" <allnor@tele.ntnu.no> wrote in news:1156521775.946503.254890@
74g2000cwt.googlegroups.com:

> I have never agreed to item 3, which I can only interpret to
> state that there is some characteristic unique to human
> words, as opposed to both non-anthropogenic sounds or
> anthropogenic sounds that are not words. 

There is huge research along such lines in the language area.  Richard 
Aslin and Elissa Newport spring to mind as possible search terms.

-- 
Scott
Reverse name to reply

Scott Seidman wrote:
> "Rune Allnor" <allnor@tele.ntnu.no> wrote in news:1156489566.536459.51370@
> 75g2000cwc.googlegroups.com:
>
> > The term "all possible" is insane from both an engineering and a R&D
> > point of view. How do you know that you have covered **all**possible**
> > variants of anything?
>
>
>
> "All possible" of course approaches hyperbole, but its a rather convenient
> term to make your goals understood in a short sentence in lay language,
> which is what's usually required if you expect to successfully attract
> funding from government sources.

Ah, now we are approaching interesting waters.

For some reason, I have noted that quite a few people who have
gained extreme levels of skills in some subject tend to regard
other people who have not gained the same levels in general,
in the same subjects in particular, as somehow "inferior" or
even all out "stupid".

I don't agree with such views.

In my experience, literate people -- even laymen what my own
field of work is concerned -- are actually able to comprehend
written text. The term "all possible" has a very definite meaning:
It means that a technique covers "all possible" variations of
whatever, and in that sense, the technique is perfect. Pepole
tend to understand such claims. What's more, they are able
to review test results and expect the method in question to
actually cover "all possible" situations.

Stay clear of projects where such claims are made. Somebody
in the funding institutions will, some day, start asking questions
why the explicitly stated goals are not met.

Rune

Scott Seidman wrote:
> "Rune Allnor" <allnor@tele.ntnu.no> wrote in news:1156521775.946503.254890@
> 74g2000cwt.googlegroups.com:
>
> > I have never agreed to item 3, which I can only interpret to
> > state that there is some characteristic unique to human
> > words, as opposed to both non-anthropogenic sounds or
> > anthropogenic sounds that are not words.
>
> There is huge research along such lines in the language area.  Richard
> Aslin and Elissa Newport spring to mind as possible search terms.

Is their test(s) able to include the tongue clicks of the Namib
bushmen?
And distinguish that from a six-year old clicking and singing to
herself
in a make-believe language while playing in the garden? Does their
test(s) distiguish between human speech and a parrot uttering the same
phrase? Does their test(s) separate human sound from synthesized
computer voices?

The claim that a technique classifies something as anthropogenic, as
opposed to parrots or computers, and word, as opposed to humming
or make-believe languages, is pretty bold.

It only takes one failed test to kick the feet away under it.

Rune

"Rune Allnor" <allnor@tele.ntnu.no> wrote in news:1156522496.169055.197210
@i42g2000cwa.googlegroups.com:

> 
> I don't agree with such views.
> 
> In my experience, literate people -- even laymen what my own
> field of work is concerned -- are actually able to comprehend
> written text. The term "all possible" has a very definite meaning:
> It means that a technique covers "all possible" variations of
> whatever, and in that sense, the technique is perfect. Pepole
> tend to understand such claims. What's more, they are able
> to review test results and expect the method in question to
> actually cover "all possible" situations.
> 
> Stay clear of projects where such claims are made. Somebody
> in the funding institutions will, some day, start asking questions
> why the explicitly stated goals are not met.
> 
> Rune
> 


I'm merely suggesting that in documents like this, you really don't know 
what the real claims are simply by reading the abstract.  If you could, 
then journals would consist of a stack of abstracts.  

-- 
Scott
Reverse name to reply