Events Calendar

PhD Preliminary Oral Exam – Xingyi Li

Leveraging Structures of the Data in Deep Learning

The performance of deep learning frameworks could be significantly improved through considering the particular underlying structures for each dataset. In the first part of this work, we summarize our two prior works about boosting deep learning models through leveraging structures of the data. In the first work, we theoretically justify that, in convolutional neural networks (CNNs), neighborhoods of a pixel should be redefined as its highest correlated spatial locations, in order to achieve a lower generalization error. Based on the correlation pattern, we propose a data-driven approach to design multiple layers of different customized filter shapes by repeatedly solving lasso problems. In the second work, we address the problem of scale-invariance in deep learning. We propose ScaleNet to predict object scales. Through recursively applying ScaleNet and rescaling, pretrained deep networks can identify objects with significantly different scales from the training set. In the last part of this work, we propose to explore PointConv based frameworks to tackle the problem of scale & rotation invariance. PointConv is a novel convolution operation that can be applied on point clouds. It takes coordinates of points as inputs to generate corresponding weights for convolution. We provide two possible configurations for this study as well as some preliminary results.

Co Advisor: Xiaoli Fern
Co Advisor: Fuxin Li
Committee: Raviv Raich
Committee: Prasad Tadepalli
GCR: Brett Tyler

Friday, May 31 at 11:00am to 1:00pm

Kelley Engineering Center, 1005
110 SW Park Terrace, Corvallis, OR 97331

Event Type

Lecture or Presentation

Event Topic

Research

Organization
Electrical Engineering and Computer Science
Contact Name

Calvin Hughes

Contact Email

calvin.hughes@oregonstate.edu

Subscribe
Google Calendar iCal Outlook

Recent Activity