<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=us-ascii">
</head>
<body style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class="">
<div class="BodyFragment"><font size="2" class=""><span style="font-size: 11pt;" class="">
<div class="PlainText">This is an announcement of Xiao Zhang's MS Presentation<br class="">
===============================================<br class="">
Candidate: Xiao Zhang<br class="">
<br class="">
Date: Friday, May 12, 2023<br class="">
<br class="">
Time: 10 am CST<br class="">
<br class="">
Location: JCL 298<br class="">
<br class="">
M.S. Paper Title: Self-Supervised visual representation learning<br class="">
<br class="">
Abstract: In this talk, I will discuss my research on self-supervised visual representation learning, which consists of two parts. The primary focus of the talk will be on the first part, where we propose a framework for learning visual representation from
hierarchical grouping. We operationalize grouping via a contour detector that partitions an image into regions, followed by merging of those regions into a tree hierarchy. A small supervised dataset suffices for training this grouping primitive. Across a large
unlabeled dataset, we apply this learned primitive to automatically predict hierarchical region structure. These predictions guide self-supervised contrastive feature learning: we task a deep network with producing per-pixel embeddings whose pairwise distances
respect the region hierarchy. Experiments demonstrate that our approach can serve as state-of-the-art generic pretraining, benefiting downstream tasks. We additionally explore applications for semantic region search and video-based object instance tracking. <br class="">
In the second part of my presentation, I will discuss two recent works on learning visual representation from a generative model. The first work employs the generative adversarial network for feature learning, utilizing proposed structural adversarial objectives.
The second work involves extracting dense representation from a denoising autoencoder using layer-distributed spectral clustering.<br class="">
<br class="">
Advisors: Michael Maire<br class="">
<br class="">
Committee Members: Michael Maire, Rebecca Willett, and Greg Shakhnarovich<br class="">
</div>
</span></font></div>
<div class="BodyFragment"><font size="2" class=""><span style="font-size: 11pt;" class=""></span></font></div>
<br class="">
</body>
</html>