[Colloquium] Shen/Dissertation Defense/Oct 19, 2017

[Margaret Jaffey via Colloquium]

       Department of Computer Science/The University of Chicago

                     *** Dissertation Defense ***


Candidate:  Jiajun Shen

Date:  Thursday, October 19, 2017

Time:  10:00 AM

Place:  Ryerson 255

Title: Transformation Invariance and Equivariance in Deep Learning

Abstract:
A major challenge for object recognition is to correctly perceive the
image objects despite the extraneous variations in the data such as
shifting, rotation, deformation, etc. It would be much easier for the
vision tasks if such task-irrelevant transformation variabilities were
removed from the data. The recent success of deep learning approaches
has its roots in the ability to build feature representations that are
invariant to variations caused by nuisance factors. The expressiveness
of deep networks allows the models to disentangle the underlying
factors of variations in the data and the training signals guide the
models to learn feature representations that are robust to
task-irrelevant variations. However, such variations need to be
observed from the training data. Otherwise, a traditional deep neural
network without special architectural design would not generalize to
these variations. To address this concern, we study the problem of
achieving transformation invariance and equivariance in deep learning.

We show how some existing approaches, such as the stacked statistical
model with rotatable features and the spatial transformer network, are
imperfect at learning feature representations that are invariant or
equivariant to transformations. To search for an alternative, we
develop a training mechanism to learn transformation-invariant feature
representations, where feature maps of canonical images are used as
soft targets to guide a deep neural network to produce the same
feature representations even when the input images are transformed. As
a result, our framework can obtain transformation-invariant feature
representations and makes it possible to take advantage of unlabeled
data that contains an enormous amount of variations. Additionally, we
seek architectural changes to the existing deep learning models and
propose a framework for training deep neural networks with optimal
instantiations. By introducing latent variables to parametrize the
transformations of the data for each class, our approach is able to
obtain the optimal instantiations while training for the downstream
tasks. Another direction we explore is the use of 3D CAD models to
render 2D images as a data augmentation approach. Rendering 2D images
from 3D models allows for a more compact way of representing an object
class and the models are able to observe more data variations during
training. Competitive experimental results are demonstrated by these
methods, and our analysis shows that they can be promising directions
to achieve transformation invariance and equivariance in deep
learning.

Jiajun's advisor is Prof. Yali Amit

Login to the Computer Science Department website for details,
including a draft copy of the dissertation:

 https://www.cs.uchicago.edu/phd/phd_announcements#jiajun

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey            margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 156)               (773) 702-6011
The University of Chicago      http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=