<html><head><meta http-equiv="Content-Type" content="text/html; charset=utf-8"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class=""><div class=""><div dir="auto" class="" style="color: rgb(0, 0, 0); caret-color: rgb(0, 0, 0); word-wrap: break-word; line-break: after-white-space;"><font face="Arial" class=""><b class="" style="font-size: 14px;"><font color="#000000" class="">University of Chicago and Toyota Technological Institute at Chicago<br class="">Machine Learning Seminar Series</font></b><br class=""><br class=""><br class=""><b class="" style="font-size: 14px;">Yali Amit</b><br class=""><span class="" style="font-size: 14px;">UChicago Statistics</span><br class=""><br class=""><b class="" style="font-size: 14px;">Wednesday, May 15, 2019 at 1:00 - 2:00 pm<br class="">Harper Center (Booth) Room 219</b><br class=""><br class=""><br class=""><b class="" style="font-size: 14px;">Title:</b></font></div><div dir="auto" class="" style="caret-color: rgb(0, 0, 0); color: rgb(0, 0, 0); word-wrap: break-word; line-break: after-white-space;"><font face="Arial" class=""><span class="" style="font-size: 11.5pt; font-family: "Times New Roman", serif;">Optimization of latent variables in deep network applications</span> <br class=""><b class="" style="font-size: 14px;"><font face="Arial" class="" style="font-weight: normal;"><b class=""><br class=""></b></font></b></font></div><div dir="auto" class="" style="caret-color: rgb(0, 0, 0); color: rgb(0, 0, 0); word-wrap: break-word; line-break: after-white-space;"><font face="Arial" class=""><span class=""><font face="Arial" class="" style="font-size: 14px;"><b class="">Abstract:</b><br class=""></font><div class=""><font color="#000000" class=""><span class="" style="font-size: 14px;">The primary use of deep networks is to provide direct prediction of response variables, whether discrete or continuous, following a forward pass through the network. In some cases however, we know that the values of certain well defined but unobserved latent variables are key to successful prediction. For example the scale or rotation of an object in an image. In deep learning the typical solution is to provide extensive (also known as augmented) training sets where the expected range of values of the latent variables is well represented. In generative modeling it is more common to define a statistical model with a distribution of the observed data conditional on some unobserved latent variables, and online optimization or simulation w.r.t to the latent variables is required when computing likelihoods. I will describe some experiments, with `deformable classifiers’ where we train a deep network together with an optimization step over predefined latent variables, and require this optimization to also be performed online during test time. I will show that this enables learning with much smaller data sets, at the cost of more intensive computation, and provides as output not just the class label but the optimal instantiation of the latent variables for the test example. I will also show some application of these ideas for training generator networks. </span></font></div><font face="Arial" class="" style="font-size: 14px;"><br class=""><b class="">Bio:</b><br class=""></font><span class=""><font color="#000000" class=""><span class="" style="font-size: 14px;">Yali Amit is Professor of Statistics and Computer Science at the University of Chicago. He fields of interest are computer vision, machine learning and computational neuroscience.</span></font><br class=""></span><div class=""><span class="" style="font-size: 14px;"><span class=""><b class="">Host: Rebecca Willett</b></span></span></div></span></font></div></div><div class=""><br class=""></div><div class=""></div></body></html>