Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc.

[Stephen José Hanson]

Barak, as usual turgid and yet clear.

Exactly correct.    I agree.  Well, mostly.

however the Jurrasic--*not antidiluvian*-- nature of citations does 
require one to *read* as you so intimated in your citation "algorithm" 
-- citations are meant to be contextual and relevant.  So it is easy to 
miss references in Cognitive Science.. as sometimes its just not clear 
what is similar to what.   SO there is a matter of judgment here that 
folks could differ on.. Math and CS are a bit more obvious,  as Calculus 
was either invented by Newton or Leibnez.   The Bernoulli brothers (not 
to be confused with the Doobie brothers)  invented probability theory 
and Laplace not the revered Bayes invented bayes law (I know.. its 
sounds wrong).

But there are phase transitions in science.. so its often hard to decide 
where the citation path starts and where it ends.   As Barak, correctly 
points out, reviews of the literature, annual reviews are good targets 
to cite, because they claim to leap-frog from the past in an 
authoritative and comprehensive way.  And so maybe you are tempted to be 
lazy and not read it.  Read it.

Mostly in academics, the stakes are so low, mostly no-one seems to care 
who invented the "one sided t-test", but someone did (Braver).    But 
with 100M$ being tossed about on a daily basis, especially when the 
algorithms are being appropriated without acknowledgement and prizes are 
given out without the entire context, Juergen;s concerns are more 
tangible.   And writing a comprehensive review  is valuable and 
generates relevant and useful discussion.    So bravo to Juergen for an 
impressive historical exegesis.   Where the problem begins, is where 
does the modern era begin--2012? 2006? 1984?, 1961?, 1946?..  an ancient 
Greek guy for Arithmetic?-- well this might be a matter of taste.    I 
think its safe to say we know when something is new because it actually 
works better (the context changed--there is a *change point* if you 
will), and the citation nuance between the present and past begins to 
melt away.

Steve


On 11/15/21 9:21 AM, Barak A. Pearlmutter wrote:
> One point of scientific propriety and writing that may be getting lost
> in the scrum here, and which has I think contributed substantially to
> the somewhat woeful state of credit assignment in the field, is the
> traditional idea of what a citation *means*.
>
> If a paper says "we use the Foo Transform (Smith, 1995)" that,
> traditionally, implies that the author has actually read Smith (1995)
> and it describes the Foo Transform as used in the work being
> presented. If the author was told that the Foo Transform was actually
> discovered by Barker (1980) but the author hasn't actually verified
> that by reading Barker (1980), then the author should NOT just cite
> Barker. If the author heard that Barker (1980) is the "right" citation
> for the Foo Transform, but they got the details of it that they're
> actually using from Smith (1995) then they're supposed to say so: "We
> use the Foo Transform as described in Smith (1995), attributed to
> Barker (1980) by someone I met in line for the toilet at NeurIPS
> 2019".
>
> This seemingly-antediluvian practice is to guard against people citing
> "Barker (1980)" as saying something that it actually doesn't say,
> proving a theorem that it doesn't, defining terms ("rate code", cough
> cough) in a fashion that is not consistent with Barker's actual
> definitions, etc. Iterated violations of this often manifest as
> repeated and successive simplification of an idea, a so-called game of
> telephone, until something not even true is sagely attributed to some
> old publication that doesn't actually say it.
>
> So if you want to cite, say, Seppo Linnainmaa for Reverse Mode
> Automatic Differentiation, you need to have actually read it yourself.
> Otherwise you need to do a bounce citation: "Linnainman (1982)
> described by Schmidhuber (2021) as exhibiting a Fortran implementation
> of Reverse Mode Automatic Differentiation" or something like that.
>
> This is also why it's considered fine to simply cite a textbook or
> survey paper: nobody could possibly mistake those as the original
> source, but they may well be where the author actually got it from.
>
> To bring this back to the present thread: I must confess that I have
> not actually read many of the old references Jürgen brings up.
> Certainly "X (1960) invented deep learning" is not enough to allow
> someone to cite them. It's not even enough for a bounce citation. What
> did they *actually* do? What is Jürgen saying they actually did?
>
-- 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20211115/d9e139a4/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.png
Type: image/png
Size: 19957 bytes
Desc: not available
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20211115/d9e139a4/attachment.png>