Connectionists: Connectionists Digest, Vol 764, Issue 1

[Juyang Weng]

Dear Suganthan,
thank you for your message about what you did in your neural network
experiments.
If I were you, I would not simply trust what advisees said to me, but I
trust my understanding of the current severe lack of generalization power
in shallow data fitting by CNNs and LSTMs (including adversarial learning),
etc.

As I stated in my Post-Selection paper in IJCNN 2021, we must demand
transparency in the Post-Selection stage, such as reporting the
distribution of performances of all networks that have been trained.
Without presentation of such a distribution, we should not simply trust a
terse statement like the one from your advisees.  Our own experiments have
shown a huge variation in such a distribution!

Why? Error-backprop in neural networks is like "Chairman Mao Zedong did
error-backprop in China".  Chairman Mao may present only one lucky case,
like a nuclear bomb, to brag about the Chinese planned economy.  For those
who like to get more an intuitive explanation, watch this YouTube video:
https://youtu.be/VpsufMtia14

The neural network community needs a fundamental cultural change: citation
integrity and transparency of all networks trained.

Best regards,
-John

On Wed, Nov 17, 2021 at 5:20 AM Ponnuthurai Nagaratnam Suganthan <
EPNSugan at ntu.edu.sg> wrote:

> Dear John,
> My understanding is that we use test set only once to determine test
> accuracy and that is the last step. We just report test results and end.
> I’m not aware of any selection afterwards. If anyone imdoes anything like
> that, that’d be cheating 😁
> Best Regards
> Suganthan
>
> ------------------------------
> *From:* Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu> on
> behalf of Juyang Weng <juyang.weng at gmail.com>
> *Sent:* Wednesday, 17 November 2021 9:30 am
> *To:* Post Connectionists <connectionists at mailman.srv.cs.cmu.edu>
> *Subject:* Re: Connectionists: Connectionists Digest, Vol 764, Issue 1
>
> Dear Juergen,
>
> I respectfully waited till people have had enough time to respond to your
> plagiarism allegations.
>
> Many people probably are not aware of a much more severe problem than the
> plagiarism you correctly raised:
>
> I would like to raise here that error-backprop is a major technical flaw
> in many types of neural networks (CNN, LSTM, etc.) buried in a protocol
> violation called Post-Selection Using Test Sets (PSUTS).
> See this IJCNN 2021 paper:
> J. Weng, "On Post Selections Using Test Sets (PSUTS) in AI", in Proc.
> International Joint Conference on Neural Networks, pp. 1-8, Shengzhen,
> China, July 18-22, 2021. PDF file
> <http://www.cse.msu.edu/~weng/research/PSUTS-IJCNN2021rvsd-cite.pdf>.
>
> Those who do not agree with me please respond.
>
> Best regards,
> -John
> ----------------------------------------------------------------------
>
> Message: 1
> Date: Sun, 14 Nov 2021 16:47:36 +0000
> From: Schmidhuber Juergen <juergen at idsia.ch>
> To: "connectionists at cs.cmu.edu" <connectionists at cs.cmu.edu>
> Subject: Re: Connectionists: Scientific Integrity, the 2021 Turing
>         Lecture, etc.
> Message-ID: <532DC982-9F4B-41F8-9AB4-AD21314C6472 at supsi.ch>
> Content-Type: text/plain; charset="utf-8"
>
> Dear all, thanks for your public comments, and many additional private
> ones!
>
> So far nobody has challenged the accuracy of any of the statements in the
> draft report currently under massive open peer review:
>
>
> https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html
>
> Nevertheless, some of the recent comments will trigger a few minor
> revisions in the near future.
>
> Here are a few answers to some of the public comments:
>
> Randall O'Reilly wrote: "I vaguely remember someone making an interesting
> case a while back that it is the *last* person to invent something that
> gets all the credit." Indeed, as I wrote in Science (2011, reference
> [NASC3] in the report): "As they say: Columbus did not become famous
> because he was the first to discover America, but because he was the last."
> Sure, some people sometimes assign the "inventor" title to the person that
> should be truly called the "popularizer." Frequently, this is precisely due
> to the popularizer packaging the work of others in such a way that it
> becomes easily digestible. But this is not to say that their receipt of the
> title is correct or that we shouldn't do our utmost to correct it; their
> receipt of such title over the ones that are actually deserving of it is
> one of the most enduring issues in scientific history.
>
> As Stephen Jos? Hanson wrote: "Well, to  popularize is not to invent. Many
> of Juergen's concerns could be solved with some scholarship, such that
> authors look sometime before 2006 for other relevant references."
>
> Randy also wrote: "Sometimes, it is not the basic equations etc that
> matter: it is the big picture vision." However, the same vision has almost
> always been there in the earlier work on neural nets. It's just that the
> work was ahead of its time. It's only in recent years that we have the
> datasets and the computational power to realize those big pictures visions.
> I think you would agree that simply scaling something up isn't the same as
> inventing it. If it were, then the name "Newton" would have little meaning
> to people nowadays.
>
> Jonathan D. Cohen wrote: " ...it is also worth noting that science is an
> *intrinsically social* endeavor, and therefore communication is a
> fundamental factor." Sure, but let?s make sure that this cannot be used as
> a justification of plagiarism! See Sec. 5 of the report.
>
> Generally speaking, if B plagiarizes A but inspires C, whom should C cite?
> The answer is clear.
>
> Ponnuthurai Nagaratnam Suganthan wrote: "The name `deep learning' came
> about recently." Not so. See references in Sec. X of the report: the
> ancient term "deep learning" (explicitly mentioned by ACM) was actually
> first introduced to Machine Learning by Dechter (1986), and to NNs by
> Aizenberg et al (2000).
>
> Tsvi Achler wrote: "Models which have true feedback (e.g. back to their
> own inputs) cannot learn by backpropagation but there is plenty of evidence
> these types of connections exist in the brain and are used during
> recognition. Thus they get ignored: no talks in universities, no featuring
> in `premier' journals and no funding. [...] Lastly Feedforward methods are
> predominant in a large part because they have financial backing from large
> companies with advertising and clout like Google and the self-driving craze
> that never fully materialized." This is very misleading - see Sec. A, B,
> and C of the report which are about recurrent nets with feedback,
> especially LSTM, heavily used by Google and others, on your smartphone
> since 2015. Recurrent NNs are general computers that can compute anything
> your laptop can compute, including any computable model with feedback "back
> to the inputs." My favorite proof from over 30 years ago: a little
> subnetwork can be used to build a NAND gate, an!
>  d a big recurrent network of NAND gates can emulate the CPU of your
> laptop. (See also answers by Dan Levine, Gary Cottrell, and Juyang Weng.)
> However, as Asim Roy pointed out, this discussion deviates from the
> original topic of improper credit assignment. Please use another thread for
> this.
>
> Randy also wrote: "Should Newton be cited instead of Rumelhart et al, for
> backprop, as Steve suggested? Seriously, most of the math powering today's
> models is just calculus and the chain rule." This is so misleading in
> several ways - see Sec. XII of the report: "Some claim that
> `backpropagation is just the chain rule of Leibniz (1676) & L'Hopital
> (1696).' No, it is the efficient way of applying the chain rule to big
> networks with differentiable nodes (there are also many inefficient ways of
> doing this). It was not published until 1970" by Seppo Linnainmaa. Of
> course, the person to cite is Linnainmaa.
>
> Randy also wrote: "how little Einstein added to what was already
> established by Lorentz and others". Juyang already respectfully objected to
> this misleading statement.
>
> I agree with what Anand Ramamoorthy wrote: "Setting aside broader aspects
> of the social quality of the scientific enterprise, let's take a look at a
> simpler thing; individual duty. Each scientist has a duty to science (as an
> intellectual discipline) and the scientific community, to uphold
> fundamental principles informing the conduct of science. Credit should be
> given wherever it is due - it is a matter of duty, not preference or
> `strategic vale' or boosting someone because they're a great populariser.
> ... Crediting those who disseminate is fine and dandy, but should be for
> those precise contributions, AND the originators of an idea/method/body of
> work ought to be recognised - this is perhaps a bit difficult when the work
> is obscured by history, but not impossible. At any rate, if one has novel
> information of pertinence w.r.t original work, then the right action is
> crystal clear."
>
> See also Sec. 5 of the report: "As emphasized earlier:[DLC][HIN] `The
> inventor of an important method should get credit for inventing it. They
> may not always be the one who popularizes it. Then the popularizer should
> get credit for popularizing it - but not for inventing it.' If one
> "re-invents" something that was already known, and only becomes aware of it
> later, one must at least clarify it later, and correctly give credit in
> follow-up papers and presentations."
>
> I also agree with what Zhaoping Li wrote: "I would find it hard to enter a
> scientific community if it is not scholarly. Each of us can do our bit to
> be scholarly, to set an example, if not a warning, to the next generation."
>
> Randy also wrote: "Outside of a paper specifically on the history of a
> field, does it really make sense to "require" everyone to cite obscure old
> papers that you can't even get a PDF of on google scholar?" This sounds
> almost like a defense of plagiarism. That's what time stamps of patents and
> papers are for. A recurring point of the report is: the awardees did not
> cite the prior art - not even in later surveys written when the true
> origins of this work were well-known.
>
> Here I fully agree with what Marina Meila wrote: "Since credit is a form
> of currency in academia, let's look at the `hard currency' rewards of
> invention. Who gets them? The first company to create a new product usually
> fails. However, the interesting thing is that society (by this I mean the
> society most of us we work in) has found it necessary to counteract this,
> and we have patent laws to protect the rights of the inventors. The point
> is not whether patent laws are effective or not, it's the social norm they
> implement. That to protect invention one should pay attention to rewarding
> the original inventors, whether we get the `product' directly from them or
> not."
>
> J?rgen
>
-- 
Juyang (John) Weng
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20211117/b5f3c6b3/attachment.html>