<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
  </head>
  <body text="#000000" bgcolor="#ecca99">
    <p><font size="+1"><font face="monospace">Great,</font></font></p>
    <p><font size="+1"><font face="monospace">these are all grammar
          strings.. nothing semantic --right?</font></font></p>
    <p><font size="+1"><font face="monospace">Gary and I had confusion
          about that.. but I read the paper..</font></font></p>
    <p><font size="+1"><font face="monospace">Steve<br>
        </font></font></p>
    <div class="moz-cite-prefix">On 6/14/22 12:11 PM, Steven T.
      Piantadosi wrote:<br>
    </div>
    <blockquote type="cite"
      cite="mid:b980a1da-8f88-f787-9e1d-b5f239c1cc1b@gmail.com">
      <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
      <p>All of our training/test data is on the github, but please let
        me know if I can help!</p>
      <p>Steve</p>
      <p><br>
      </p>
      <div class="moz-cite-prefix">On 6/13/22 06:13, Gary Marcus wrote:<br>
      </div>
      <blockquote type="cite"
        cite="mid:9D553D9E-CA9A-43E6-840F-0DD05F3C4D9E@nyu.edu">
        <meta http-equiv="content-type" content="text/html;
          charset=UTF-8">
        <div dir="ltr"> I do remember the work :) Just generally
          Transformers seem more effective; a careful comparison between
          Y&P, Transformers, and your RNN approach, looking at
          generalization to novel words, would indeed be interesting.</div>
        <div dir="ltr">Cheers, </div>
        <div dir="ltr">Gary </div>
        <div dir="ltr"><br>
          <blockquote type="cite">On Jun 13, 2022, at 06:09, <a
              class="moz-txt-link-abbreviated"
              href="mailto:jose@rubic.rutgers.edu"
              moz-do-not-send="true">jose@rubic.rutgers.edu</a> wrote:<br>
            <br>
          </blockquote>
        </div>
        <blockquote type="cite">
          <div dir="ltr">
            <meta http-equiv="Content-Type" content="text/html;
              charset=UTF-8">
            <p><font size="+1"><font face="monospace">I was thinking
                  more like an RNN similar to work we had done in the
                  2000s.. on syntax.</font></font></p>
            <p><font size="+1"><font face="monospace">Stephen José
                  Hanson, Michiro Negishi; On the Emergence of Rules in
                  Neural Networks. Neural Comput 2002; 14 (9):
                  2245–2268. doi: <a class="moz-txt-link-freetext"
href="https://urldefense.com/v3/__https://doi.org/10.1162/089976602320264079__;!!BhJSzQqDqA!WCsRlT1zpBKD3ai8Ov_I79iH_HCdTlAMymGIe2ZsIdTnfZawzlMQNGZWisMjmcLBgH6SbBUZ6rtr_exEspS4Igo$"
                    moz-do-not-send="true">https://doi.org/10.1162/089976602320264079</a></font></font></p>
            <p><font size="+1"><font face="monospace">Abstract<br>
                  A simple associationist neural network learns to
                  factor abstract rules (i.e., grammars) from sequences
                  of arbitrary input symbols by inventing abstract
                  representations that accommodate unseen symbol sets as
                  well as unseen but similar grammars. The neural
                  network is shown to have the ability to transfer
                  grammatical knowledge to both new symbol vocabularies
                  and new grammars. Analysis of the state-space shows
                  that the network learns generalized abstract
                  structures of the input and is not simply memorizing
                  the input strings. These representations are context
                  sensitive, hierarchical, and based on the state
                  variable of the finite-state machines that the neural
                  network has learned. Generalization to new symbol sets
                  or grammars arises from the spatial nature of the
                  internal representations used by the network, allowing
                  new symbol sets to be encoded close to symbol sets
                  that have already been learned in the hidden unit
                  space of the network. The results are counter to the
                  arguments that learning algorithms based on weight
                  adaptation after each exemplar presentation (such as
                  the long term potentiation found in the mammalian
                  nervous system) cannot in principle extract symbolic
                  knowledge from positive examples as prescribed by
                  prevailing human linguistic theory and evolutionary
                  psychology.<br>
                </font></font><br>
            </p>
            <div class="moz-cite-prefix">On 6/13/22 8:55 AM, Gary Marcus
              wrote:<br>
            </div>
            <blockquote type="cite"
              cite="mid:60B7CD76-2DCE-4DB9-9ECD-96D046916BB8@nyu.edu">
              <meta http-equiv="content-type" content="text/html;
                charset=UTF-8">
              <div dir="ltr">– agree with Steve this is an interesting
                paper, and replicating it with a neural net would be
                interesting; cc’ing Steve Piantosi. </div>
              <div dir="ltr">— why not use a Transformer, though?</div>
              <div dir="ltr">- it is however importantly missing
                semantics. (Steve P. tells me there is some related work
                that is worth looking into). Y&P speaks to an old
                tradition of formal language work by Gold and others
                that is quite popular but IMHO misguided, because it
                focuses purely on syntax rather than semantics.  Gold’s
                work definitely motivates learnability but I have never
                taken it to seriously as a real model of language</div>
              <div dir="ltr">- doing what Y&P try to do with a rich
                artificial language that is focused around
                syntax-semantic mappings could be very interesting</div>
              <div dir="ltr">- on a somewhat but not entirely analogous
                note, i think that  the next step in vision is really
                scene understanding. We have techniques for doing object
                labeling reasonably well, but still struggle wit parts
                and wholes are important, and with relations more
                generally, which is to say we need the semantics of
                scenes. is the chair on the floor, or floating in the
                air? is it supporting the pillow? etc. is the hand a
                part of the body? is the glove a part of the body? etc</div>
              <div dir="ltr"><br>
              </div>
              <div dir="ltr">Best,</div>
              <div dir="ltr">Gary</div>
              <div dir="ltr"><br>
              </div>
              <div dir="ltr"><br>
              </div>
              <div dir="ltr"><br>
                <blockquote type="cite">On Jun 13, 2022, at 05:18, <a
                    class="moz-txt-link-abbreviated
                    moz-txt-link-freetext"
                    href="mailto:jose@rubic.rutgers.edu"
                    moz-do-not-send="true">jose@rubic.rutgers.edu</a>
                  wrote:<br>
                  <br>
                </blockquote>
              </div>
              <blockquote type="cite">
                <div dir="ltr">
                  <p><font size="+1">Again, I think a relevant project
                      here  would be to attempt to replicate with
                      DL-rnn, Yang and Piatiadosi's PNAS language
                      learning system--which is a completely symbolic--
                      and very general over the Chomsky-Miller grammer
                      classes.   Let me know, happy to collaborate on
                      something like this.<br>
                    </font></p>
                  <p><font size="+1">Best</font></p>
                  <p><font size="+1">Steve<br>
                    </font></p>
                </div>
              </blockquote>
            </blockquote>
          </div>
        </blockquote>
      </blockquote>
    </blockquote>
  </body>
</html>