<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
  </head>
  <body text="#000000" bgcolor="#ecca99">
    <p><font size="+1">+1</font></p>
    <p><font size="+1">So I was the 4th program chair of NIPS back in
        1991, the 5th General Chair.<br>
      </font></p>
    <p><font size="+1">I have been to every advisory/exec board meeting
        since that time and almost every NIPS till Covid hit--last one
        in 2019.</font></p>
    <p><font size="+1">I have never seen or experienced a "cabal" of
        Terry "cohorts", just the opposite.   Terry has maintained a
        high integrity, honest environment and listened to all input and
        concerns.     NIPS is truly a democratic, serious and fair
        enterprise and organization, which is due to Terry's careful and
        light touch on the rudder.<br>
      </font></p>
    <p><font size="+1">Sue's points are obvious to anyone who has
        curated this conference.  Its really impossible to "subtly" or
        otherwise change the direction of the conference or prevent good
        research from being accepted.  Is there a huge MISS rate.. no
        doubt.  But the conference has been a success, because of its
        transparency and its scientific diversity.</font></p>
    <p><font size="+1">I really don't understand where this is coming
        from, but certainly not from the well documented concerns that
        Juergen has raised.  He and I disagree on historical
        interpretation.. but I don't think this should be taken as
        evidence of some larger paranoid view of the field and the
        invisible hand that is controlling it.</font></p>
    <p><font size="+1">Steve<br>
      </font></p>
    <p><font size="+1"><br>
      </font></p>
    <div class="moz-cite-prefix">On 1/4/22 10:47 AM, Sue Becker wrote:<br>
    </div>
    <blockquote type="cite"
      cite="mid:14ffb6b0bc0902bb19bc1fc27b42eef8@mcmaster.ca">Pierre, 
      hI'm responding to your comment here:
      <br>
      <br>
      <blockquote type="cite">Terry: ... you have made sure, year after
        year,  that you and your BHL/CIFAR
        <br>
        friends were able to control and subtly manipulate NIPS/NeurIPS
        <br>
        (misleading the field in wrong directions, preventing news ideas
        and
        <br>
        outsiders from flourishing, and distorting credit attribution).
        <br>
        <br>
        Can you please explain to this mailing list how this serves as
        being "a
        <br>
        good role model" (to use your own words) for the next
        generation?
        <br>
      </blockquote>
      <br>
      As loathe as I am to wade into what has become a cesspool of a
      debate,  you have gone way outside the bounds of accuracy, not to
      mention civility and decency, in directing your mudslinging at
      Terry Sejnowski. If anything,  Terry deserves recognition and
      thanks for his many years of service to this community.
      <br>
      <br>
      If you think that NeurIPS is run by a bunch of insiders, try
      stepping up and volunteering your service to this conference, be a
      longtime committed reviewer, then become an Area Chair, do an
      outstanding job and be selected as the next program chair and then
      general chair. That is one path to influencing the future of the
      conference. Much more importantly, the hundreds of dedicated
      reviewers are the ones who actually determine the content of the
      meeting, by identifying the very best papers out of the thousands
      of submissions received each year.  There is no top-down control
      or manipulation over that process.
      <br>
      <br>
      Cheers,
      <br>
      Sue
      <br>
      <br>
      ---
      <br>
      Sue Becker, Professor
      <br>
      Neurotechnology and Neuroplasticity Lab, PI
      <br>
      Dept. of Psychology Neuroscience & Behaviour, McMaster
      University
      <br>
      <a class="moz-txt-link-abbreviated" href="http://www.science.mcmaster.ca/pnb/department/becker">www.science.mcmaster.ca/pnb/department/becker</a>
      <br>
      <br>
      <br>
      On 2022-01-03 09:55, Baldi,Pierre wrote:
      <br>
      <blockquote type="cite">Terry:
        <br>
        <br>
        We can all agree on the importance of mentoring the next
        generation.
        <br>
        However, given that:
        <br>
        <br>
        1) you have been in full and sole control of the NIPS/NeurIPS
        foundation
        <br>
        since the 1980s;
        <br>
        <br>
        2) you have been in full and sole control of Neural Computation
        since
        <br>
        the 1980s;
        <br>
        <br>
        3) you have extensively published in Neural Computation (and now
        also PNAS);
        <br>
        <br>
        4) you have made sure, year after year,  that you and your
        BHL/CIFAR
        <br>
        friends were able to control and subtly manipulate NIPS/NeurIPS
        <br>
        (misleading the field in wrong directions, preventing news ideas
        and
        <br>
        outsiders from flourishing, and distorting credit attribution).
        <br>
        <br>
        Can you please explain to this mailing list how this serves as
        being "a
        <br>
        good role model" (to use your own words) for the next
        generation?
        <br>
        <br>
        Or did you mean it in a more cynical way--indeed this is one of
        the
        <br>
        possible ways for a scientist to be "successful"?
        <br>
        <br>
        --Pierre
        <br>
        <br>
        <br>
        <br>
        On 1/2/2022 12:29 PM, Terry Sejnowski wrote:
        <br>
        <blockquote type="cite">We would be remiss not to acknowledge
          that backprop would not be
          <br>
          possible without the calculus,
          <br>
          so Isaac newton should also have been given credit, at least
          as much
          <br>
          credit as Gauss.
          <br>
          <br>
          All these threads will be sorted out by historians one hundred
          years
          <br>
          from now.
          <br>
          Our precious time is better spent moving the field forward. 
          There is
          <br>
          much more to discover.
          <br>
          <br>
          A new generation with better computational and mathematical
          tools than
          <br>
          we had back
          <br>
          in the last century have joined us, so let us be good role
          models and
          <br>
          mentors to them.
          <br>
          <br>
          Terry
          <br>
          <br>
          -----
          <br>
          <br>
          On 1/2/2022 5:43 AM, Schmidhuber Juergen wrote:
          <br>
          <blockquote type="cite">Asim wrote: "In fairness to Jeffrey
            Hinton, he did acknowledge the
            <br>
            work of Amari in a debate about connectionism at the ICNN’97
            .... He
            <br>
            literally said 'Amari invented back propagation'..." when he
            sat next
            <br>
            to Amari and Werbos. Later, however, he failed to cite
            Amari’s
            <br>
            stochastic gradient descent (SGD) for multilayer NNs
            (1967-68)
            <br>
            [GD1-2a] in his 2015 survey [DL3], his 2021 ACM lecture
            [DL3a], and
            <br>
            other surveys.  Furthermore, SGD [STO51-52] (Robbins, Monro,
            Kiefer,
            <br>
            Wolfowitz, 1951-52) is not even backprop. Backprop is just a
            <br>
            particularly efficient way of computing gradients in
            differentiable
            <br>
            networks, known as the reverse mode of automatic
            differentiation, due
            <br>
            to Linnainmaa (1970) [BP1] (see also Kelley's precursor of
            1960
            <br>
            [BPa]). Hinton did not cite these papers either, and in 2019
            <br>
            embarrassingly did not hesitate to accept an award for
            having
            <br>
            "created ... the backpropagation algorithm” [HIN]. All
            references and
            <br>
            more on this can be found in the report, especially in !
            <br>
          </blockquote>
          Se!
          <br>
          <blockquote type="cite">  c. XII.
            <br>
            <br>
            The deontology of science requires: If one "re-invents"
            something
            <br>
            that was already known, and only becomes aware of it later,
            one must
            <br>
            at least clarify it later [DLC], and correctly give credit
            in all
            <br>
            follow-up papers and presentations. Also, ACM's Code of
            Ethics and
            <br>
            Professional Conduct [ACM18] states: "Computing
            professionals should
            <br>
            therefore credit the creators of ideas, inventions, work,
            and
            <br>
            artifacts, and respect copyrights, patents, trade secrets,
            license
            <br>
            agreements, and other methods of protecting authors' works."
            LBH didn't.
            <br>
            <br>
            Steve still doesn't believe that linear regression of 200
            years ago
            <br>
            is equivalent to linear NNs. In a mature field such as math
            we would
            <br>
            not have such a discussion. The math is clear. And even
            today, many
            <br>
            students are taught NNs like this: let's start with a linear
            <br>
            single-layer NN (activation = sum of weighted inputs). Now
            minimize
            <br>
            mean squared error on the training set. That's good old
            linear
            <br>
            regression (method of least squares). Now let's introduce
            multiple
            <br>
            layers and nonlinear but differentiable activation
            functions, and
            <br>
            derive backprop for deeper nets in 1960-70 style (still used
            today,
            <br>
            half a century later).
            <br>
            <br>
            Sure, an important new variation of the 1950s (emphasized by
            Steve)
            <br>
            was to transform linear NNs into binary classifiers with
            threshold
            <br>
            functions. Nevertheless, the first adaptive NNs (still
            widely used
            <br>
            today) are 1.5 centuries older except for the name.
            <br>
            <br>
            Happy New Year!
            <br>
            <br>
            Jürgen
            <br>
            <br>
            <br>
            <blockquote type="cite">On 2 Jan 2022, at 03:43, Asim Roy
              <a class="moz-txt-link-rfc2396E" href="mailto:ASIM.ROY@asu.edu"><ASIM.ROY@asu.edu></a> wrote:
              <br>
              <br>
              And, by the way, Paul Werbos was also there at the same
              debate. And
              <br>
              so was Teuvo Kohonen.
              <br>
              <br>
              Asim
              <br>
              <br>
              -----Original Message-----
              <br>
              From: Asim Roy
              <br>
              Sent: Saturday, January 1, 2022 3:19 PM
              <br>
              To: Schmidhuber Juergen <a class="moz-txt-link-rfc2396E" href="mailto:juergen@idsia.ch"><juergen@idsia.ch></a>;
              <a class="moz-txt-link-abbreviated" href="mailto:connectionists@cs.cmu.edu">connectionists@cs.cmu.edu</a>
              <br>
              Subject: RE: Connectionists: Scientific Integrity, the
              2021 Turing
              <br>
              Lecture, etc.
              <br>
              <br>
              In fairness to Jeffrey Hinton, he did acknowledge the work
              of Amari
              <br>
              in a debate about connectionism at the ICNN’97
              (International
              <br>
              Conference on Neural Networks) in Houston. He literally
              said "Amari
              <br>
              invented back propagation" and Amari was sitting next to
              him. I
              <br>
              still have a recording of that debate.
              <br>
              <br>
              Asim Roy
              <br>
              Professor, Information Systems
              <br>
              Arizona State University
              <br>
              <a class="moz-txt-link-freetext" href="https://isearch.asu.edu/profile/9973">https://isearch.asu.edu/profile/9973</a>
              <br>
              <a class="moz-txt-link-freetext" href="https://lifeboat.com/ex/bios.asim.roy">https://lifeboat.com/ex/bios.asim.roy</a>
              <br>
            </blockquote>
            <br>
            On 2 Jan 2022, at 02:31, Stephen José Hanson
            <a class="moz-txt-link-rfc2396E" href="mailto:jose@rubic.rutgers.edu"><jose@rubic.rutgers.edu></a>
            <br>
            wrote:
            <br>
            <br>
            Juergen:  Happy New Year!
            <br>
            <br>
            "are not quite the same"..
            <br>
            <br>
            I understand that its expedient sometimes to use linear
            regression to
            <br>
            approximate the Perceptron.(i've had other connectionist
            friends tell
            <br>
            me the same thing) which has its own incremental update
            rule..that is
            <br>
            doing <0,1> classification.    So I guess if you don't
            like the
            <br>
            analogy to logistic regression.. maybe Fisher's LDA?  This
            whole
            <br>
            thing still doesn't scan for me.
            <br>
            <br>
            So, again the point here is context.   Do you really believe
            that
            <br>
            Frank Rosenblatt didn't reference Gauss/Legendre/Laplace
            because it
            <br>
            slipped his mind??   He certainly understood modern
            statistics (of
            <br>
            the 1940s and 1950s)
            <br>
            <br>
            Certainly you'd agree that FR could have referenced linear
            regression
            <br>
            as a precursor, or "pretty similar" to what he was working
            on, it
            <br>
            seems disingenuous to imply he was plagiarizing Gauss et
            al.--right? 
            <br>
            Why would he?
            <br>
            <br>
            Finally then, in any historical reconstruction, I can think
            of, it
            <br>
            just doesn't make sense.    Sorry.
            <br>
            <br>
            Steve
            <br>
            <br>
            <br>
            <blockquote type="cite">-----Original Message-----
              <br>
              From: Connectionists
              <a class="moz-txt-link-rfc2396E" href="mailto:connectionists-bounces@mailman.srv.cs.cmu.edu"><connectionists-bounces@mailman.srv.cs.cmu.edu></a>
              <br>
              On Behalf Of Schmidhuber Juergen
              <br>
              Sent: Friday, December 31, 2021 11:00 AM
              <br>
              To: <a class="moz-txt-link-abbreviated" href="mailto:connectionists@cs.cmu.edu">connectionists@cs.cmu.edu</a>
              <br>
              Subject: Re: Connectionists: Scientific Integrity, the
              2021 Turing
              <br>
              Lecture, etc.
              <br>
              <br>
              Sure, Steve, perceptron/Adaline/other similar methods of
              the
              <br>
              1950s/60s are not quite the same, but the obvious origin
              and
              <br>
              ancestor of all those single-layer  “shallow learning”
              <br>
              architectures/methods is indeed linear regression; today’s
              simplest
              <br>
              NNs minimizing mean squared error are exactly what they
              had 2
              <br>
              centuries ago. And the first working deep learning methods
              of the
              <br>
              1960s did NOT really require “modern” backprop (published
              in 1970 by
              <br>
              Linnainmaa [BP1-5]). For example, Ivakhnenko & Lapa
              (1965) [DEEP1-2]
              <br>
              incrementally trained and pruned their deep networks layer
              by layer
              <br>
              to learn internal representations, using regression and a
              separate
              <br>
              validation set. Amari (1967-68)[GD1] used stochastic
              gradient
              <br>
              descent [STO51-52] to learn internal representations
              WITHOUT
              <br>
              “modern" backprop in his multilayer perceptrons. Jürgen
              <br>
              <br>
              <br>
              <blockquote type="cite">On 31 Dec 2021, at 18:24, Stephen
                José Hanson
                <br>
                <a class="moz-txt-link-rfc2396E" href="mailto:jose@rubic.rutgers.edu"><jose@rubic.rutgers.edu></a> wrote:
                <br>
                <br>
                Well the perceptron is closer to logistic regression...
                but the
                <br>
                heaviside function  of course is <0,1>   so
                technically not related
                <br>
                to linear regression which is using covariance to
                estimate betas...
                <br>
                <br>
                does that matter?  Yes, if you want to be hyper
                correct--as this
                <br>
                appears to be-- Berkson (1944) coined the logit.. as log
                odds.. for
                <br>
                probabilistic classification.. this was formally
                developed by Cox
                <br>
                in the early 60s, so unlikely even in this case to be a
                precursor
                <br>
                to perceptron.
                <br>
                <br>
                My point was that DL requires both Learning algorithm
                (BP) and an
                <br>
                architecture.. which seems to me much more responsible
                for the the
                <br>
                success of Dl.
                <br>
                <br>
                S
                <br>
                <br>
                <br>
                <br>
                On 12/31/21 4:03 AM, Schmidhuber Juergen wrote:
                <br>
                <blockquote type="cite">Steve, this is not about machine
                  learning in general, just about deep
                  <br>
                  learning vs shallow learning. However, I added the
                  Pandemonium -
                  <br>
                  thanks for that! You ask: how is a linear regressor of
                  1800
                  <br>
                  (Gauss/Legendre) related to a linear neural network?
                  It's formally
                  <br>
                  equivalent, of course! (The only difference is that
                  the weights are
                  <br>
                  often called beta_i rather than w_i.) Shallow
                  learning: one adaptive
                  <br>
                  layer. Deep learning: many adaptive layers. Cheers,
                  Jürgen
                  <br>
                  <br>
                  <br>
                  <br>
                  <br>
                  <blockquote type="cite">On 31 Dec 2021, at 00:28,
                    Stephen José Hanson
                    <br>
                    <a class="moz-txt-link-rfc2396E" href="mailto:jose@rubic.rutgers.edu"><jose@rubic.rutgers.edu></a>
                    <br>
                    wrote:
                    <br>
                    <br>
                    Despite the comprehensive feel of this it still
                    appears to me to
                    <br>
                    be  too focused on Back-propagation per se.. (except
                    for that
                    <br>
                    pesky Gauss/Legendre ref--which still baffles me at
                    least how
                    <br>
                    this is related to a "neural network"), and at the
                    same time it
                    <br>
                    appears to be missing other more general
                    epoch-conceptually
                    <br>
                    relevant cases, say:
                    <br>
                    <br>
                    Oliver Selfridge  and his Pandemonium model.. which
                    was a
                    <br>
                    hierarchical feature analysis system.. which
                    certainly was in the
                    <br>
                    air during the Neural network learning heyday...in
                    fact, Minsky
                    <br>
                    cites Selfridge as one of his mentors.
                    <br>
                    <br>
                    Arthur Samuels:  Checker playing system.. which
                    learned a
                    <br>
                    evaluation function from a hierarchical search.
                    <br>
                    <br>
                    Rosenblatt's advisor was Egon Brunswick.. who was a
                    gestalt
                    <br>
                    perceptual psychologist who introduced the concept
                    that the world
                    <br>
                    was stochastic and the the organism had to adapt to
                    this variance
                    <br>
                    somehow.. he called it "probabilistic
                    functionalism"  which
                    <br>
                    brought attention to learning, perception and
                    decision theory,
                    <br>
                    certainly all piece parts of what we call neural
                    networks.
                    <br>
                    <br>
                    There are many other such examples that influenced
                    or provided
                    <br>
                    context for the yeasty mix that was 1940s and 1950s
                    where Neural
                    <br>
                    Networks  first appeared partly due to PItts and
                    McCulloch which
                    <br>
                    entangled the human brain with computation and early
                    computers
                    <br>
                    themselves.
                    <br>
                    <br>
                    I just don't see this as didactic, in the sense of a
                    conceptual
                    <br>
                    view of the  multidimensional history of the        
                    field, as
                    <br>
                    opposed to  a 1-dimensional exegesis of mathematical
                    threads
                    <br>
                    through various statistical algorithms.
                    <br>
                    <br>
                    Steve
                    <br>
                    <br>
                    On 12/30/21 1:03 PM, Schmidhuber Juergen wrote:
                    <br>
                    <br>
                    <blockquote type="cite">Dear connectionists,
                      <br>
                      <br>
                      in the wake of massive open online peer review,
                      public comments
                      <br>
                      on the connectionists mailing list [CONN21] and
                      many additional
                      <br>
                      private comments (some by well-known deep learning
                      pioneers)
                      <br>
                      helped to update and improve upon version 1 of the
                      report. The
                      <br>
                      essential statements of the text remain unchanged
                      as their
                      <br>
                      accuracy remains unchallenged. I'd like to thank
                      everyone from
                      <br>
                      the bottom of my heart for their feedback up until
                      this point
                      <br>
                      and hope everyone will be satisfied with the
                      changes. Here is
                      <br>
                      the revised version 2 with over 300 references:
                      <br>
                      <br>
                      <br>
                      <br>
<a class="moz-txt-link-freetext" href="https://urldefense.com/v3/__https://people.idsia.ch/*juergen/scient">https://urldefense.com/v3/__https://people.idsia.ch/*juergen/scient</a>
                      <br>
ific-integrity-turing-award-deep-learning.html__;fg!!IKRxdwAv5BmarQ
                      <br>
!NsJ4lf4yO2BDIBzlUVfGKvTtf_QXY8dpZaHzCSzHCvEhXGJUTyRTzZybDQg-DZY$
                      <br>
                      <br>
                      <br>
                      <br>
                      In particular, Sec. II has become a brief history
                      of deep
                      <br>
                      learning up to the 1970s:
                      <br>
                      <br>
                      Some of the most powerful NN architectures (i.e.,
                      recurrent NNs)
                      <br>
                      were discussed in 1943 by McCulloch and Pitts
                      [MC43] and
                      <br>
                      formally analyzed in 1956 by Kleene [K56] - the
                      closely related
                      <br>
                      prior work in physics by Lenz, Ising, Kramers, and
                      Wannier dates
                      <br>
                      back to the 1920s [L20][I25][K41][W45]. In 1948,
                      Turing wrote up
                      <br>
                      ideas related to artificial evolution [TUR1] and
                      learning NNs.
                      <br>
                      He failed to formally publish his ideas though,
                      which explains
                      <br>
                      the obscurity of his thoughts here. Minsky's
                      simple neural SNARC
                      <br>
                      computer dates back to 1951. Rosenblatt's
                      perceptron with a
                      <br>
                      single adaptive layer learned in 1958 [R58]
                      (Joseph [R61]
                      <br>
                      mentions an earlier perceptron-like device by
                      Farley & Clark);
                      <br>
                      Widrow & Hoff's similar Adaline learned in
                      1962 [WID62]. Such
                      <br>
                      single-layer "shallow learning" actually started
                      around 1800
                      <br>
                      when Gauss & Legendre introduced linear
                      regression and the
                      <br>
                      method of least squares [DL1-2] - a famous early
                      example of
                      <br>
                      pattern recognition and generalization from
                      training!
                      <br>
                    </blockquote>
                  </blockquote>
                </blockquote>
              </blockquote>
            </blockquote>
          </blockquote>
           !
          <br>
          <blockquote type="cite">  d!
            <br>
            <blockquote type="cite">at!
              <br>
              <blockquote type="cite">
                <blockquote type="cite">a through a parameterized
                  predictor is Gauss' rediscovery of the
                  <br>
                  asteroid Ceres based on previous astronomical
                  observations. Deeper
                  <br>
                  multilayer perceptrons (MLPs) were discussed by
                  Steinbuch
                  <br>
                  [ST61-95] (1961), Joseph [R61] (1961), and Rosenblatt
                  [R62]
                  <br>
                  (1962), who wrote about "back-propagating errors" in
                  an MLP with a
                  <br>
                  hidden layer [R62], but did not yet have a general
                  deep learning
                  <br>
                  algorithm for deep MLPs  (what's now called
                  backpropagation is
                  <br>
                  quite different and was first published by Linnainmaa
                  in 1970
                  <br>
                  [BP1-BP5][BPA-C]). Successful learning in deep
                  architectures
                  <br>
                  started in 1965 when Ivakhnenko & Lapa published
                  the first
                  <br>
                  general, working learning algorithms for deep MLPs
                  with
                  <br>
                  arbitrarily many hidden layers (already containing the
                  now popular
                  <br>
                  multiplicative gates) [DEEP1-2][DL1-2]. A paper of
                  1971 [DEEP2]
                  <br>
                  already described a deep learning net with 8 layers,
                  trained by
                  <br>
                  their highly cited method which was still popular in
                  the new
                  <br>
                  millennium [DL2], especially in Eastern Europ!
                  <br>
                </blockquote>
              </blockquote>
            </blockquote>
          </blockquote>
          e!
          <br>
          <blockquote type="cite">
            <blockquote type="cite">, w!
              <br>
              <blockquote type="cite">
                <blockquote type="cite">here much of Machine Learning
                  was born [MIR](Sec. 1)[R8]. LBH !
                  <br>
                  failed to
                  <br>
                  cite this, just like they failed to cite Amari [GD1],
                  who in 1967
                  <br>
                  proposed stochastic gradient descent [STO51-52] (SGD)
                  for MLPs and
                  <br>
                  whose implementation [GD2,GD2a] (with Saito) learned
                  internal
                  <br>
                  representations at a time when compute was billions of
                  times more
                  <br>
                  expensive than today (see also Tsypkin's work
                  [GDa-b]). (In 1972,
                  <br>
                  Amari also published what was later sometimes called
                  the Hopfield
                  <br>
                  network or Amari-Hopfield Network [AMH1-3].)
                  Fukushima's now
                  <br>
                  widely used deep convolutional NN architecture was
                  first
                  <br>
                  introduced in the 1970s [CNN1].
                  <br>
                  <br>
                  <blockquote type="cite">
                    <blockquote type="cite">Jürgen
                      <br>
                      <br>
                      <br>
                      <br>
                      <br>
                      ******************************
                      <br>
                      <br>
                      On 27 Oct 2021, at 10:52, Schmidhuber Juergen
                      <br>
                      <br>
                      <a class="moz-txt-link-rfc2396E" href="mailto:juergen@idsia.ch"><juergen@idsia.ch></a>
                      <br>
                      <br>
                      wrote:
                      <br>
                      <br>
                      Hi, fellow artificial neural network enthusiasts!
                      <br>
                      <br>
                      The connectionists mailing list is perhaps the
                      oldest mailing
                      <br>
                      list on ANNs, and many neural net pioneers are
                      still subscribed
                      <br>
                      to it. I am hoping that some of them - as well as
                      their
                      <br>
                      contemporaries - might be able to provide
                      additional valuable
                      <br>
                      insights into the history of the field.
                      <br>
                      <br>
                      Following the great success of massive open online
                      peer review
                      <br>
                      (MOOR) for my 2015 survey of deep learning (now
                      the most cited
                      <br>
                      article ever published in the journal Neural
                      Networks), I've
                      <br>
                      decided to put forward another piece for MOOR. I
                      want to thank the
                      <br>
                      many experts who have already provided me with
                      comments on it.
                      <br>
                      Please send additional relevant references and
                      suggestions for
                      <br>
                      improvements for the following draft directly to
                      me at
                      <br>
                      <br>
                      <a class="moz-txt-link-abbreviated" href="mailto:juergen@idsia.ch">juergen@idsia.ch</a>
                      <br>
                      <br>
                      :
                      <br>
                      <br>
                      <br>
                      <br>
<a class="moz-txt-link-freetext" href="https://urldefense.com/v3/__https://people.idsia.ch/*juergen/scient">https://urldefense.com/v3/__https://people.idsia.ch/*juergen/scient</a>
                      <br>
ific-integrity-turing-award-deep-learning.html__;fg!!IKRxdwAv5BmarQ
                      <br>
!NsJ4lf4yO2BDIBzlUVfGKvTtf_QXY8dpZaHzCSzHCvEhXGJUTyRTzZybDQg-DZY$
                      <br>
                      <br>
                      <br>
                      <br>
                      The above is a point-for-point critique of factual
                      errors in
                      <br>
                      ACM's justification of the ACM A. M. Turing Award
                      for deep
                      <br>
                      learning and a critique of the Turing Lecture
                      published by ACM
                      <br>
                      in July 2021. This work can also be seen as a
                      short history of
                      <br>
                      deep learning, at least as far as ACM's errors and
                      the Turing
                      <br>
                      Lecture are concerned.
                      <br>
                      <br>
                      I know that some view this as a controversial
                      topic. However, it
                      <br>
                      is the very nature of science to resolve
                      controversies through
                      <br>
                      facts. Credit assignment is as core to scientific
                      history as it
                      <br>
                      is to machine learning. My aim is to ensure that
                      the true
                      <br>
                      history of our field is preserved for posterity.
                      <br>
                      <br>
                      Thank you all in advance for your help!
                      <br>
                      <br>
                      Jürgen Schmidhuber
                      <br>
                      <br>
                      <br>
                      <br>
                      <br>
                      <br>
                      <br>
                    </blockquote>
                    --
                    <br>
                    <signature.png>
                    <br>
                    <br>
                  </blockquote>
                </blockquote>
                --
                <br>
                <signature.png>
                <br>
              </blockquote>
              <br>
            </blockquote>
            <br>
          </blockquote>
          <br>
          <br>
        </blockquote>
      </blockquote>
    </blockquote>
    <div class="moz-signature">-- <br>
      <img src="cid:part1.72A5F2FE.079C852E@rubic.rutgers.edu"
        border="0"></div>
  </body>
</html>