[Theory] REMINDER: 1/26 Talks at TTIC: Tolga Birdal, Stanford University

Mary Marre mmarre at ttic.edu
Mon Jan 25 16:02:06 CST 2021


*When:*      Tuesday, January 26th at* 11:10 am CT*



*Where:*     Zoom Virtual Talk (*register in advance here
<https://uchicagogroup.zoom.us/webinar/register/WN_KJ4Af7uDQqadp5OgFOS6OQ>*)



*Who: *       Tolga Birdal, Stanford University


*Title: Non-Euclidean Machine Learning for 3D Computer Vision*

*Abstract*: We understand the world by interacting with the objects and
agents we observe. This Kantian empirical realism called *experience* is
made possible by the a priori Euclidean constraints on space. While being
subject to limits of scales and tolerances of our senses, such a *flat* view
of the world has been the driving force in many engineering fields laying
the foundations of the first AI systems. For example, the typical data that
learning systems consume such as text, audio, or images are organized into a
 *grid/lattice* just like the pixel matrix underlying an image. This makes
the processing easy and allows researchers to devise domain specific
algorithms. On the other hand, the typical output of a learning machine is
a set of real valued numbers that best explain a downstream task such as
predicting the weather temperature. Neural networks as the de-facto choices
are then responsible for optimally mapping the space of the input to the
output space, both of which are *assumed Euclidean*. Unfortunately, for a
variety of applications ranging from 3D computer vision to analysis of
social networks or protein structures, the assumptions posed by Euclidean
geometry cease to hold. For instance, *3D surfaces*, *point clouds*, *trees*
 or *graphs* are types of inputs whose data points neither satisfy the
Pythagorean theorem nor can be arranged into a grid without information
loss. Furthermore, entities such as *rotations*, *permutations* (orderings)
or *probability distributions* cannot be regressed (or predicted) without
restricting the domain of real numbers. In other words, they lie on a lower
dimensional *sub-manifold* endowed with a certain geometric structure.

In my research, I challenge this Euclidean perspective and propose to work
on the non-Euclidean, curved structure of the environments that surround us.
 I coin such mapping of non-Euclidean input to non-Euclidean output the
*non-Euclidean machine learning*. My ultimate goal is to arm the autonomous
systems running on 3D data with capabilities of scene or object-level
reasoning natively on the *geometric* nature of the 3D perception problem.
In this talk, I will summarize my previous research on processing 3D point
clouds to enable understanding of rigid and non-rigid dynamics. I will also
investigate how to provide the additional *uncertainty* information for the
problems at hand. Finally, I will open a window into the future
technologies and applications such approach can enable.

*BIO*: Tolga Birdal is a Postdoctoral Research Fellow at Stanford
University. He carries his research within the Geometric Computing Group of
Prof. Leonidas Guibas. Previously, Tolga has defended his PhD thesis at the
Computer Vision Group, Chair for Computer Aided Medical Procedures,
Technical University of Munich led by Prof. Nassir Navab. He was also a
Doktorand at Siemens AG. Tolga completed his Bachelors as an Electronics
Engineer at Sabanci University in 2008. In his subsequent postgraduate
programme, he studied Computational Science and Engineering at Technical
University of Munich. In continuation to his Master's thesis on “3D
Deformable Surface Recovery Using RGBD Cameras”, he focused his research
and development on large object detection, pose estimation and
reconstruction using point clouds. Tolga is awarded both Ernst von Siemens
Scholarship and EMVA Young Professional Award 2016 for his PhD work. He has
several publications at the well respected venues such as NeurIPS, CVPR,
ICCV, ECCV, IROS, ICASSP and 3DV. Aside from his academic life, Tolga is a
natural Entrepreneur. He has co-founded multiple companies including
Befunky, a widely used web based image processing platform. For further
information, visit tbirdal.me, https://profiles.stanford.edu/tolga-birdal.

*Host:* Matthew Walter <mwalter at ttic.edu>


Mary C. Marre
Faculty Administrative Support
*Toyota Technological Institute*
*6045 S. Kenwood Avenue*
*Room 517*
*Chicago, IL  60637*
*p:(773) 834-1757*
*f: (773) 357-6970*
*mmarre at ttic.edu <mmarre at ttic.edu>*


On Thu, Jan 21, 2021 at 2:55 PM Mary Marre <mmarre at ttic.edu> wrote:

> *When:*      Tuesday, January 26th at* 11:10 am CT*
>
>
>
> *Where:*     Zoom Virtual Talk (*register in advance here
> <https://uchicagogroup.zoom.us/webinar/register/WN_KJ4Af7uDQqadp5OgFOS6OQ>*
> )
>
>
>
> *Who: *       Tolga Birdal, Stanford University
>
>
> *Title: Non-Euclidean Machine Learning for 3D Computer Vision*
>
> *Abstract*: We understand the world by interacting with the objects and
> agents we observe. This Kantian empirical realism called *experience* is
> made possible by the a priori Euclidean constraints on space. While being
> subject to limits of scales and tolerances of our senses, such a *flat* view
> of the world has been the driving force in many engineering fields laying
> the foundations of the first AI systems. For example, the typical data
> that learning systems consume such as text, audio, or images are
> organized into a *grid/lattice* just like the pixel matrix underlying an
> image. This makes the processing easy and allows researchers to devise
> domain specific algorithms. On the other hand, the typical output of a
> learning machine is a set of real valued numbers that best explain a
> downstream task such as predicting the weather temperature. Neural networks
> as the de-facto choices are then responsible for optimally mapping the
> space of the input to the output space, both of which are *assumed
> Euclidean*. Unfortunately, for a variety of applications ranging from 3D
> computer vision to analysis of social networks or protein structures, the
> assumptions posed by Euclidean geometry cease to hold. For instance, *3D
> surfaces*, *point clouds*, *trees* or *graphs* are types of inputs whose
> data points neither satisfy the Pythagorean theorem nor can be arranged
> into a grid without information loss. Furthermore, entities such as
> *rotations*, *permutations* (orderings) or *probability distributions* cannot
> be regressed (or predicted) without restricting the domain of real numbers.
> In other words, they lie on a lower dimensional *sub-manifold* endowed
> with a certain geometric structure.
>
> In my research, I challenge this Euclidean perspective and propose to work
> on the non-Euclidean, curved structure of the environments that surround
> us. I coin such mapping of non-Euclidean input to non-Euclidean output the
>  *non-Euclidean machine learning*. My ultimate goal is to arm the
> autonomous systems running on 3D data with capabilities of scene or
> object-level reasoning natively on the *geometric* nature of the 3D
> perception problem. In this talk, I will summarize my previous research on
> processing 3D point clouds to enable understanding of rigid and non-rigid
> dynamics. I will also investigate how to provide the additional
> *uncertainty* information for the problems at hand. Finally, I will open
> a window into the future technologies and applications such approach can
> enable.
>
> *BIO*: Tolga Birdal is a Postdoctoral Research Fellow at Stanford
> University. He carries his research within the Geometric Computing Group of
> Prof. Leonidas Guibas. Previously, Tolga has defended his PhD thesis at the
> Computer Vision Group, Chair for Computer Aided Medical Procedures,
> Technical University of Munich led by Prof. Nassir Navab. He was also a
> Doktorand at Siemens AG. Tolga completed his Bachelors as an Electronics
> Engineer at Sabanci University in 2008. In his subsequent postgraduate
> programme, he studied Computational Science and Engineering at Technical
> University of Munich. In continuation to his Master's thesis on “3D
> Deformable Surface Recovery Using RGBD Cameras”, he focused his research
> and development on large object detection, pose estimation and
> reconstruction using point clouds. Tolga is awarded both Ernst von Siemens
> Scholarship and EMVA Young Professional Award 2016 for his PhD work. He has
> several publications at the well respected venues such as NeurIPS, CVPR,
> ICCV, ECCV, IROS, ICASSP and 3DV. Aside from his academic life, Tolga is a
> natural Entrepreneur. He has co-founded multiple companies including
> Befunky, a widely used web based image processing platform. For further
> information, visit tbirdal.me, https://profiles.stanford.edu/tolga-birdal.
>
> *Host:* Matthew Walter <mwalter at ttic.edu>
>
>
> Mary C. Marre
> Faculty Administrative Support
> *Toyota Technological Institute*
> *6045 S. Kenwood Avenue*
> *Room 517*
> *Chicago, IL  60637*
> *p:(773) 834-1757*
> *f: (773) 357-6970*
> *mmarre at ttic.edu <mmarre at ttic.edu>*
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/theory/attachments/20210125/731ad728/attachment-0001.html>


More information about the Theory mailing list