November 2, 2019 Conference paper Open Access

Nguyen, Toan Q.; Salazar, Julian

MARC21 XML Export

<?xml version='1.0' encoding='UTF-8'?>
<record xmlns="http://www.loc.gov/MARC21/slim">
  <leader>00000nam##2200000uu#4500</leader>
  <datafield tag="041" ind1=" " ind2=" ">
    <subfield code="a">eng</subfield>
  </datafield>
  <controlfield tag="005">20200120174454.0</controlfield>
  <controlfield tag="001">3525484</controlfield>
  <datafield tag="700" ind1=" " ind2=" ">
    <subfield code="u">Amazon AWS AI</subfield>
    <subfield code="a">Salazar, Julian</subfield>
  </datafield>
  <datafield tag="856" ind1="4" ind2=" ">
    <subfield code="s">345932</subfield>
    <subfield code="z">md5:5a6c18ef21719ddacdab79deec9a4b39</subfield>
    <subfield code="u">https://zenodo.org/record/3525484/files/IWSLT2019_paper_26.pdf</subfield>
  </datafield>
  <datafield tag="542" ind1=" " ind2=" ">
    <subfield code="l">open</subfield>
  </datafield>
  <datafield tag="260" ind1=" " ind2=" ">
    <subfield code="c">2019-11-02</subfield>
  </datafield>
  <datafield tag="909" ind1="C" ind2="O">
    <subfield code="p">openaire</subfield>
    <subfield code="p">user-iwslt2019</subfield>
    <subfield code="o">oai:zenodo.org:3525484</subfield>
  </datafield>
  <datafield tag="100" ind1=" " ind2=" ">
    <subfield code="u">University of Notre Dame</subfield>
    <subfield code="a">Nguyen, Toan Q.</subfield>
  </datafield>
  <datafield tag="245" ind1=" " ind2=" ">
    <subfield code="a">Transformers without Tears: Improving the Normalization of Self-Attention</subfield>
  </datafield>
  <datafield tag="980" ind1=" " ind2=" ">
    <subfield code="a">user-iwslt2019</subfield>
  </datafield>
  <datafield tag="540" ind1=" " ind2=" ">
    <subfield code="u">https://creativecommons.org/licenses/by/4.0/legalcode</subfield>
    <subfield code="a">Creative Commons Attribution 4.0 International</subfield>
  </datafield>
  <datafield tag="650" ind1="1" ind2="7">
    <subfield code="a">cc-by</subfield>
    <subfield code="2">opendefinition.org</subfield>
  </datafield>
  <datafield tag="520" ind1=" " ind2=" ">
    <subfield code="a">&lt;p&gt;We evaluate three simple, normalization-centric changes to improve Transformer training. First, we show that pre-norm residual connections (PRENORM) and smaller initializations enable warmup-free, validation-based training with large learning rates. Second, we propose&amp;nbsp;l2&amp;nbsp;normalization with a single scale parameter (SCALENORM) for faster training and better performance. Finally, we reaffirm the effectiveness of normalizing word embeddings to a fixed length (FIXNORM). On five low-resource translation pairs from TED Talks-based corpora, these changes always converge, giving an average +1.1 BLEU over state-of-the-art bilingual baselines and a new 32.8 BLEU on IWSLT &amp;#39;15 English-Vietnamese. We ob- serve sharper performance curves, more consistent gradient norms, and a linear relationship between activation scaling and decoder depth. Surprisingly, in the high-resource setting (WMT &amp;#39;14 English-German), SCALENORM&amp;nbsp;and FIXNORM&amp;nbsp;remain competitive but PRENORM&amp;nbsp;degrades performance.&lt;/p&gt;</subfield>
  </datafield>
  <datafield tag="773" ind1=" " ind2=" ">
    <subfield code="n">doi</subfield>
    <subfield code="i">isVersionOf</subfield>
    <subfield code="a">10.5281/zenodo.3525483</subfield>
  </datafield>
  <datafield tag="024" ind1=" " ind2=" ">
    <subfield code="a">10.5281/zenodo.3525484</subfield>
    <subfield code="2">doi</subfield>
  </datafield>
  <datafield tag="980" ind1=" " ind2=" ">
    <subfield code="a">publication</subfield>
    <subfield code="b">conferencepaper</subfield>
  </datafield>
</record>