<html><body style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><head><meta http-equiv="Content-Type" content="text/html; charset=us-ascii"></head><br class=""><span id="docs-internal-guid-de668218-7fff-220a-42c4-c3d01efd7aaf" class="" style="font-family: -webkit-standard;"><div class="" style="line-height: 1.38; margin-top: 0pt; margin-bottom: 0pt;"></div><div class="" style="line-height: 1.38; margin-top: 0pt; margin-bottom: 0pt;"><br class=""></div></span><div dir="auto" class="" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><div dir="auto" class="" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><div dir="auto" class="" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><div dir="auto" class="" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><div class="" style="orphans: 2; widows: 2;"><span class="" style="font-size: 14px;"></span></div><div class="" style="orphans: 2; widows: 2;"></div></div></div></div></div><div dir="auto" class="" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><div dir="auto" class="" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><div dir="auto" class="" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><div dir="auto" class="" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;"><div class="" style="orphans: 2; widows: 2;"><div class="" style="margin: 0in 0in 0.0001pt;"><b class=""><font size="4" class="">Chi Jin</font></b></div><div class="" style="margin: 0in 0in 0.0001pt;"><span class="" style="font-size: 14px;"><i class="">UC Berkeley<br class=""></i></span><div class=""><span class="" style="font-size: 14px;"><i class=""><span class="Apple-tab-span" style="white-space: pre;"> </span></i></span></div></div><div class="" style="margin: 0in 0in 0.0001pt;"><b class=""><font class="" style="font-size: 14px;"><br class=""></font></b></div><div class="" style="margin: 0in 0in 0.0001pt;"><span class=""><b class=""><font class="" size="4">Thursday, February 21, 2019 at 2:00 pm<br class="">Crerar 390</font></b><br class=""></span></div></div><div class="" style="orphans: 2; widows: 2;"><span class="" style="font-size: 14px;"><br class=""></span></div><div class=""><br class=""></div><div class=""><div class=""><font class="" style="font-size: 15px;"><span class=""><b class="" style="color: rgb(33, 33, 33);">Title: <span class="Apple-tab-span" style="white-space: pre;"> </span></b></span></font><font color="#212121" class=""><span class="" style="font-size: 15px;">Why Do Simple Algorithms Work So Well?</span></font></div><div class="" style="color: rgb(33, 33, 33);"><b class=""><font class="" style="font-size: 15px;"><br class=""></font></b></div><div class="" style="color: rgb(33, 33, 33);"><b class=""><font class="" style="font-size: 15px;">Abstract:</font></b></div><div class=""><font class=""><span class=""><div class="" style="font-variant-ligatures: normal; background-color: rgb(255, 255, 255);"><div class=""><font class=""><span class=""><span id="docs-internal-guid-bf351367-7fff-5067-8c8e-0130cc54a81c" class=""><div class="" style="line-height: 1.38; margin-top: 0pt; margin-bottom: 0pt;"><span class=""><span class=""><span class=""><span style="font-size: 15px; white-space: pre-wrap;" class="">While state-of-the-art machine learning models are deep, large-scale, sequential and highly nonconvex, the backbone of modern learning algorithms are simple algorithms such as stochastic gradient descent, or Q-learning (in the case of reinforcement learning tasks). A basic question endures---why do simple algorithms work so well even in these challenging settings? <br class=""><br class="">This talk focuses on two fundamental problems: (1) in nonconvex optimization, can gradient descent escape saddle points efficiently? (2) in reinforcement learning, is Q-learning sample efficient? We will provide the first line of provably positive answers to both questions. In particular, we will show that simple modifications to these classical algorithms guarantee significantly better properties, which explains the underlying mechanisms behind their favorable performance in practice.</span><br class=""></span><span class="" style="font-size: 15px; white-space: pre-wrap;"> </span><span class="" style="font-size: 15px; white-space: pre-wrap;"><br class=""></span><b class="" style="font-size: 15px; white-space: pre-wrap;">Bio:</b><br class=""></span><span class=""><span style="font-size: 15px; white-space: pre-wrap;" class="">Chi Jin is a Ph.D. candidate in Computer Science at UC Berkeley, advised by Michael I. Jordan. He received a B.S. in Physics from Peking University. His research interests lie in machine learning, statistics, and optimization, with his PhD work primarily focused on nonconvex optimization and reinforcement learning. </span><br class=""><span style="font-size: 15px; white-space: pre-wrap;" class=""> </span></span></span></div><div class="" style="line-height: 1.38; margin-top: 0pt; margin-bottom: 0pt;"><span class=""><i class="" style="font-size: 15px; color: rgb(34, 34, 34);"><b class="">Host: Rebecca Willett</b></i></span></div><div class="" style="line-height: 1.38; margin-top: 0pt; margin-bottom: 0pt;"><span class=""><i class="" style="font-size: 15px; color: rgb(34, 34, 34);"><b class=""><br class=""></b></i></span></div><div class="" style="line-height: 1.38; margin-top: 0pt; margin-bottom: 0pt;"><span class=""><i class="" style="font-size: 15px; color: rgb(34, 34, 34);"><b class="">PDF: </b></i></span></div><div class="" style="line-height: 1.38; margin-top: 0pt; margin-bottom: 0pt;"><span class=""><i class="" style="font-size: 15px; color: rgb(34, 34, 34);"><b class=""></b></i></span></div></span></span></font></div></div></span></font></div></div></div></div></div></div></body></html>