<html>
  <head>

    <meta http-equiv="content-type" content="text/html; charset=UTF-8">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <br>
    <div class="moz-forward-container"><br>
      <div dir="ltr">A talk of possible interest<br>
        <br>
        <div class="gmail_quote">
          <div dir="ltr" class="gmail_attr">---------- Forwarded message
            ---------<br>
            From: <strong class="gmail_sendername" dir="auto">Gina M.
              Holland</strong> <span dir="ltr">&lt;<a
                href="mailto:gholland@princeton.edu"
                moz-do-not-send="true">gholland@princeton.edu</a>&gt;</span><br>
            Date: Wed, 27 Feb 2019 at 11:28<br>
            Subject: NEXT: PACM COLLOQUIUM - Monday, March 4 - Shay
            Moran (Princeton University)<br>
            To: <a href="mailto:pacmdistribution@math.princeton.edu"
              moz-do-not-send="true">pacmdistribution@math.princeton.edu</a>
            &lt;<a href="mailto:pacmdistribution@math.princeton.edu"
              moz-do-not-send="true">pacmdistribution@math.princeton.edu</a>&gt;<br>
          </div>
          <br>
          <br>
          <div link="blue" vlink="purple" lang="EN-US">
            <div class="m_4967748544967758751WordSection1">
              <p class="MsoNormal"><span
                  style="font-size:12.0pt;font-family:&quot;Times New
                  Roman&quot;,serif">The next PACM Colloquium will be
                  given by
                  <b>Shay Moran (Princeton University)</b></span><b><span
                    style="font-size:9.0pt;font-family:&quot;Times New
                    Roman&quot;,serif"></span></b></p>
              <p class="MsoNormal"><span
                  style="font-size:12.0pt;font-family:&quot;Times New
                  Roman&quot;,serif"><br>
                  =================<br>
                  <b>DATE/LOCATION: Monday, March 4 2019, 4:00 – 5:00 PM
                    / 214 Fine Hall</b><br>
                  <br>
                  SPEAKER: <b>Shay Moran (Princeton University)</b></span><b><span
                    style="font-size:9.0pt;font-family:&quot;Times New
                    Roman&quot;,serif"></span></b></p>
              <p class="MsoNormal"><span
                  style="font-size:12.0pt;font-family:&quot;Times New
                  Roman&quot;,serif"><br>
                  TITLE:  </span><b><span
                    style="font-size:12.0pt;font-family:&quot;Times New
                    Roman&quot;,serif">The Optimal Approximation Factor
                    in Density Estimation</span></b><span
                  style="font-size:12.0pt;font-family:&quot;Times New
                  Roman&quot;,serif"></span></p>
              <p class="MsoNormal"><span
                  style="font-size:12.0pt;font-family:&quot;Times New
                  Roman&quot;,serif"> </span></p>
              <p class="MsoNormal"><span
                  style="font-size:12.0pt;font-family:&quot;Times New
                  Roman&quot;,serif">ABSTRACT:</span> </p>
              <p class="MsoNormal"> </p>
              <div style="border:none;border-bottom:double windowtext
                2.25pt;padding:0in 0in 1.0pt 0in">
                <p class="MsoNormal"><span
                    style="font-size:12.0pt;font-family:&quot;Times New
                    Roman&quot;,serif">Consider the following problem:
                    given arbitrary densities q1,q2 and a sample-access
                    to an unknown target density p, find which of the
                    qi's is closer to p in the total variation distance.
                  </span></p>
                <p class="MsoNormal"><span
                    style="font-size:8.0pt;font-family:&quot;Times New
                    Roman&quot;,serif"> </span></p>
                <p class="MsoNormal"><span
                    style="font-size:12.0pt;font-family:&quot;Times New
                    Roman&quot;,serif">A beautiful (and simple) result
                    due to Yatracos (1985) shows that this problem is
                    tractable in the following sense: there exists an
                    algorithm that uses O(epsilon^{-2}) samples from p
                    and outputs qi such that with high
                    probability, TV(qi,p) &lt;= 3*OPT + epsilon, where
                    OPT= min{TV(q1,p),TV(q2,p)}. Moreover, this result
                    extends to any finite class of densities: there
                    exists an algorithm that outputs the best density in
                    Q up to a multiplicative approximation factor of 3.</span></p>
                <p class="MsoNormal"><span
                    style="font-size:8.0pt;font-family:&quot;Times New
                    Roman&quot;,serif"> </span></p>
                <p class="MsoNormal"><span
                    style="font-size:12.0pt;font-family:&quot;Times New
                    Roman&quot;,serif">We complement and extend this
                    result by showing that: (i) the factor 3 cannot be
                    improved if one restricts the algorithm to output a
                    density from Q, and (ii) if one allows the algorithm
                    to output arbitrary densities (e.g. a mixture of
                    densities from Q), then the approximation factor can
                    be reduced to 2, which is optimal. In particular
                    this demonstrates an advantage of improper learning
                    over proper in this setup.</span></p>
                <p class="MsoNormal"><span
                    style="font-size:8.0pt;font-family:&quot;Times New
                    Roman&quot;,serif"> </span></p>
                <p class="MsoNormal"><span
                    style="font-size:12.0pt;font-family:&quot;Times New
                    Roman&quot;,serif">Our  algorithms rely on
                    estimating carefully chosen surrogates metrics to
                    the total variation, and our sample complexity
                    bounds exploit techniques from Adaptive Data
                    Analysis.</span></p>
                <p class="MsoNormal"><span
                    style="font-size:12.0pt;font-family:&quot;Times New
                    Roman&quot;,serif">Joint work with Olivier Bousquet
                    (Google brain) and Daniel Kane (UCSD).</span></p>
                <p class="MsoNormal"><i><span
                      style="font-size:12.0pt;font-family:&quot;Times
                      New Roman&quot;,serif;color:#1f497d"> </span></i></p>
                <p class="MsoNormal"><i><span
                      style="font-size:12.0pt;font-family:&quot;Times
                      New Roman&quot;,serif;color:#7f7f7f">Shay Moran is
                      a Postdoctoral fellow at the Computer Science
                      Department in Princeton University. He graduated
                      from the Technion in September ’16. During 2017 he
                      was a postdoctoral fellow at UCSD and at the
                      Simons Institute in Berkeley. During 2018 he was a
                      member at the Institute for Advanced Study. In
                      October ’19 he will join the Math Department at
                      the Technion as an assistant Professor. Shay’s
                      research interests revolves around mathematical
                      problems that arise in computer science, with a
                      focus on combinatorial problems related to machine
                      learning.</span></i></p>
                <p class="MsoNormal"><i> </i></p>
              </div>
              <p class="MsoNormal"><i><span style="color:#1f497d"> </span></i></p>
              <p class="MsoNormal"><span style="color:#1f497d"><img
                    style="width:1.0in;height:1.0in"
                    id="m_4967748544967758751Picture_x0020_1"
                    src="cid:part4.EF3808C6.5D1CE910@math.ias.edu"
                    alt="Macintosh HD:Users:theresa:Desktop:PACM
                    logo:FINAL LOGO:PACM-logo.png" class="" width="96"
                    height="96"></span></p>
              <p class="MsoNormal"><span style="color:#1f497d"><img
                    style="width:1.3333in;height:.5in"
                    id="m_4967748544967758751Picture_x0020_2"
                    src="cid:part5.499A4B18.51F59690@math.ias.edu"
                    alt="Macintosh
HD:Users:theresa:Downloads:PUsig2-pkg:Links:Black-and-white:Black:PUsig2-bw-bs.pdf"
                    class="" width="128" height="48"></span></p>
              <p class="MsoNormal"><span style="color:#1f497d"> </span></p>
              <p class="MsoNormal"> </p>
              <p class="MsoNormal"><span style="color:#1f497d"> </span></p>
            </div>
          </div>
        </div>
      </div>
    </div>
  </body>
</html>