- Lisa_teatalk - mailman.iro.umontreal.ca

Tea Talk Tomorrow at 13h30 (NIPS oral practise talk)
by Aaron Courville 22 Nov '11

22 Nov '11

Greeting people of LISA, Tomorrow we have Yann Dauphin practising his NIPS oral. Be there. NOTE: new time. When: 13h30 Wednesday Nov. 23. Where: TBA -- will let you know when I find out (hoping for AA3195). The title is "The Manifold Tangent Classifier". Maybe you could also mention it's a practice talk for NIPS. Abstract: We combine three important ideas present in previous work for building classifiers: the semi-supervised hypothesis (the input distribution contains information about the classifier), the unsupervised manifold hypothesis (data density concentrates near low-dimensional manifolds), and the manifold hypothesis for classification (different classes correspond to disjoint manifolds separated by low density). We exploit a new algorithm for capturing manifold structure (high-order contractive autoencoders) and we show how it builds a topological atlas of charts, each chart being characterized by the principal singular vectors of the Jacobian of a representation mapping. This representation learning algorithm can be stacked to yield a deep architecture, and we combine it with a domain knowledge-free version of the TangentProp algorithm to encourage the classifier to be insensitive to local directions changes along the manifold. Record-breaking results are obtained and we find that the learned tangent directions are very meaningful. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Tomorrow!
by Aaron Courville 15 Nov '11

15 Nov '11

Hey Gang, This week we have Guillaume Desjardins talking about a recent ICML paper on enhanced gradient methods. When: Nov 16th 14h00 Where: LISA Lab (AA3256) Abstract: In this tea talk, I will present recent work by KyungHyun Cho on "enhanced gradient" for RBMs. The motivation for this new gradient are two-fold. First, it is easy to show that the typical maximum likelihood gradient on the weights is a function of the gradients on the biases. Second, the RBM is over-parametrized in that multiple (visible/hidden states, parameter) configurations can lead to the same energy function. The enhanced gradient addresses both of these problems by being invariant to these bit-flip transformations. This results in faster convergence, less "dead" filters and an invariance to the actual binary representation of the data (e.g. ability to learn bit-flipped MNIST). We shall also discuss links to the natural gradient and time allowing, discuss their learning rate adaptation schedule. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Tomorrow!
by Aaron Courville 08 Nov '11

08 Nov '11

Greeting, This week we will have Olivier D. tell us about sum-product networks and his recent contributions to their theoretic development. This work will be appearing at the upcoming NIPS. When: Wednesday Nov. 9th. 14h00 Where: LISA LAB (AA3256) Title: On Sum-Product Networks Abstract: I will explain what are sum-product networks, why they can be useful, and why we care about deep ones. I will provide theoretical results on the expressive power of deep sum-product networks, showing they are able to represent some functions (exponentially) more compactly than their shallow counterpart. These results contribute to motivate recent research involving learning of deep sum-product networks, and more generally motivate research in Deep Learning. This is joint work with Yoshua. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

2 1

Reminder: Tea Talk This Morning!
by Aaron Courville 04 Nov '11

04 Nov '11

Reminder: we have a tea talk this morning at 11h30 in the LISA Lab. Details below. Cheers, Aaron On Tue, Nov 1, 2011 at 6:12 PM, Aaron Courville <aaron.courville(a)gmail.com>wrote: > Due to a conflicting Ubisoft meeting we have rescheduled Ian's Tea Talk > for Friday at 11h30. Original message, including the abstract, below. > > i.e. > New Date/Time: Friday Nov. 4th / 11h30 > > Cheers, > Aaron > > > > On Tue, Nov 1, 2011 at 6:20 AM, Aaron Courville <aaron.courville(a)gmail.com > > wrote: > >> Due to travel by some of us to a DARPA-related meeting. The tea talk this >> week has been moved to Friday. Hope you all can make it -- this talk should >> be of interest to everyone and particularly for the newer students that may >> not have had much previous exposure to statistical learning theory. The >> talk will be given by Ian Goodfellow. >> >> Date/Time: Friday Nov. 4th / 14h00 >> Location: LISA Lab (AA3256) >> >> Title: Statistical Learning Theory >> >> Abstract: >> I will give a tutorial on statistical learning theory. First I will >> show how to prove some simple probabilistic guarantees on the >> generalization error of a classifier in the PAC learning framework >> using Chebyshev's inequality. Next I will discuss how to extend these >> bounds to more realistic settings using the concept of >> Vapnik-Chervonenkis dimension. Finally I will compare structural risk >> minimization, a model selection principle based on statistical >> learning theory, to other methods of obtaining low generalization >> error. >> >> >> Cheers, >> Aaron >> >> -- >> Aaron C. Courville >> Département d’Informatique et >> de recherche opérationnelle >> Université de Montréal >> email:Aaron.Courville@gmail.com >> >> >> > > > -- > Aaron C. Courville > Département d’Informatique et > de recherche opérationnelle > Université de Montréal > email:Aaron.Courville@gmail.com > > > -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Friday, NEW TIME!
by Aaron Courville 01 Nov '11

01 Nov '11

Due to a conflicting Ubisoft meeting we have rescheduled Ian's Tea Talk for Friday at 11h30. Original message, including the abstract, below. i.e. New Date/Time: Friday Nov. 4th / 11h30 Cheers, Aaron On Tue, Nov 1, 2011 at 6:20 AM, Aaron Courville <aaron.courville(a)gmail.com>wrote: > Due to travel by some of us to a DARPA-related meeting. The tea talk this > week has been moved to Friday. Hope you all can make it -- this talk should > be of interest to everyone and particularly for the newer students that may > not have had much previous exposure to statistical learning theory. The > talk will be given by Ian Goodfellow. > > Date/Time: Friday Nov. 4th / 14h00 > Location: LISA Lab (AA3256) > > Title: Statistical Learning Theory > > Abstract: > I will give a tutorial on statistical learning theory. First I will > show how to prove some simple probabilistic guarantees on the > generalization error of a classifier in the PAC learning framework > using Chebyshev's inequality. Next I will discuss how to extend these > bounds to more realistic settings using the concept of > Vapnik-Chervonenkis dimension. Finally I will compare structural risk > minimization, a model selection principle based on statistical > learning theory, to other methods of obtaining low generalization > error. > > > Cheers, > Aaron > > -- > Aaron C. Courville > Département d’Informatique et > de recherche opérationnelle > Université de Montréal > email:Aaron.Courville@gmail.com > > > -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Friday
by Aaron Courville 01 Nov '11

01 Nov '11

Due to travel by some of us to a DARPA-related meeting. The tea talk this week has been moved to Friday. Hope you all can make it -- this talk should be of interest to everyone and particularly for the newer students that may not have had much previous exposure to statistical learning theory. The talk will be given by Ian Goodfellow. Date/Time: Friday Nov. 4th / 14h00 Location: LISA Lab (AA3256) Title: Statistical Learning Theory Abstract: I will give a tutorial on statistical learning theory. First I will show how to prove some simple probabilistic guarantees on the generalization error of a classifier in the PAC learning framework using Chebyshev's inequality. Next I will discuss how to extend these bounds to more realistic settings using the concept of Vapnik-Chervonenkis dimension. Finally I will compare structural risk minimization, a model selection principle based on statistical learning theory, to other methods of obtaining low generalization error. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Reminder: Tea Talk Today!
by Aaron Courville 19 Oct '11

19 Oct '11

Details below. On Tue, Oct 18, 2011 at 10:08 AM, Aaron Courville <aaron.courville(a)gmail.com > wrote: > Hey, > > Tomorrow David Warde-Farley will tell us about git and why we should all > love and respect it. > > Time and Date: 14h00, Wednesday Oct. 19th > Location: LISA lab (AA3256) > > Title: Collaborative Software Development With Git > > Abstract: > I will give a tutorial on the Git version control system, its particular > strengths and unique properties, as well as the most common commands and > usage patterns. We will cover interaction with local repositories, syncing > with remote repositories (including a brief branching, merging, rebasing) > and an overview of the workflow we have adopted for Theano development. A > basic knowledge of some version control system (CVS, Subversion, Mercurial) > is assumed. > > Cheers, > Aaron > > -- > Aaron C. Courville > Département d’Informatique et > de recherche opérationnelle > Université de Montréal > email:Aaron.Courville@gmail.com > > > -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Tomorrow!
by Aaron Courville 18 Oct '11

18 Oct '11

Hey, Tomorrow David Warde-Farley will tell us about git and why we should all love and respect it. Time and Date: 14h00, Wednesday Oct. 19th Location: LISA lab (AA3256) Title: Collaborative Software Development With Git Abstract: I will give a tutorial on the Git version control system, its particular strengths and unique properties, as well as the most common commands and usage patterns. We will cover interaction with local repositories, syncing with remote repositories (including a brief branching, merging, rebasing) and an overview of the workflow we have adopted for Theano development. A basic knowledge of some version control system (CVS, Subversion, Mercurial) is assumed. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

TEA TALK WEDNESDAY (NOTE NEW TIME!)
by Aaron Courville 12 Oct '11

12 Oct '11

Hi All, This week we will have a short tea talk by Myriam, Frédéric and Steven. See you there. Time and Date: 10h30-11h00 Wednesday, Oct 12th. Location: LISA lab (AA3256) Title: Common Code Workflow Abstract: Discussion on the new operating processes for code development at the lab as well as for Ubisoft's related projects. This is an opportunity to acquire a valuable experience which could make a difference on your CV. Duration: 30 minutes Your presence is required. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

2 1

Tea Talk Tomorrow!
by Aaron Courville 27 Sep '11

27 Sep '11

Greeting, This week we will have Ian Goodfellow reviewing everyone's favorite single-layer undirected graphical model: the RBM! Hope to see you there. When: 14h00, Wednesday, Sept. 28th (NOTE: NEW DAY) Where LISA LAB (AA3256) Title: RBM Boot Camp Abstract: By request from the new students in the lab, Ian will give a broad overview of RBMs. The main purpose of the talk will be to describe basic principles of structured probabilistic models, what RBMs are, why they are useful as generative models, what kinds of RBMs have been made, and the theory of how to train them. Owing to limited time, we will not cover the use of RBMs in deep learning or practical implementation details in this talk. The goal is to build a detailed yet intuitive understanding of RBMs in order to make the deep learning literature more accessible. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

LISA WANTS YOU to give a Tea Talk
by Aaron Courville 20 Sep '11

20 Sep '11

Ok, it's a new term and it's time for me to ask you to sign up for your Tea Talk. What's a Tea Talk? These are usually informal talks that can be about anything from a crazy idea you've been kicking around and want to bounce off some of your colleagues to a great paper you just read and would like to tell the lab members about. When are the Tea Talks? Tea talks are usually on Thursdays at 14h00. ** All students are expected to give at least one tea talk a year. Sign up now to reserve your spot (bonus marks for signing up for the slot this Thursday!) ** How do I sign up? Either email me with your desired time slot (some Thursday at 14h00), or preferably edit the tea talk slot directly on the LISA google calendar to include your name (i.e. change "tea talks" to "Tea Talk: Joe Shmoe"). It's as easy as that. I look forward to hearing from all of you. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 1

Reminder: Tea Talk Tomorrow
by Aaron Courville 09 Sep '11

09 Sep '11

Hi Gang, This week we will have a tea talk by Christopher Pal and Pascal Vincent about their recent vision work. Note the change in time and day. Hope to see you all there. Date & Time: Friday Sept 9th. 15h30 (It might start a little late as it will be immediately after Guillaume's predoc oral exam). Place: AA3195 (NOTE LOCATION CHANGE!) Title: "Monitoring user states during gameplay" Cameras, microphones, and heart-rate monitors have been installed to record a player's attitude during Ubisoft gametests. We will describe the computer vision processing pipeline and various machine learning approaches that we have so far explored for the sake of analysing the face-facing camera data. One goal of this work was to allow automatic detection of interesting events (that is, mainly to detect when the user's facial expression displays a state other than neutral). This is achieved with clustering and outlier detection techniques and has been implemented in a Playtest data browser, which will, among other things, greatly facilitate any future state/emotion labeling efforts on this data. We will also give a brief account of the various approaches that we explored for learning to classifiy facial emotions based on publicly available labeled datasets. Finally the presentation will open to a broader reflection on the potential for such technologies to be incorporated in the future, not only to facilitate playtest analysis, but also in original live game ideas. Note: this is a practice for a presentation that will be given at Ubisoft. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Friday
by Aaron Courville 07 Sep '11

07 Sep '11

Hi Gang, This week we will have a tea talk by Christopher Pal and Pascal Vincent about their recent vision work. Note the change in time and day. Hope to see you all there. Date & Time: Friday Sept 9th. 15h30 Place: TBA (likely the LISA lab) Title: "Monitoring user states during gameplay" Cameras, microphones, and heart-rate monitors have been installed to record a player's attitude during Ubisoft gametests. We will describe the computer vision processing pipeline and various machine learning approaches that we have so far explored for the sake of analysing the face-facing camera data. One goal of this work was to allow automatic detection of interesting events (that is, mainly to detect when the user's facial expression displays a state other than neutral). This is achieved with clustering and outlier detection techniques and has been implemented in a Playtest data browser, which will, among other things, greatly facilitate any future state/emotion labeling efforts on this data. We will also give a brief account of the various approaches that we explored for learning to classifiy facial emotions based on publicly available labeled datasets. Finally the presentation will open to a broader reflection on the potential for such technologies to be incorporated in the future, not only to facilitate playtest analysis, but also in original live game ideas. Note: this is a practice for a presentation that will be given at Ubisoft. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

REMINDER: Tea Talk Today!
by Aaron Courville 25 Aoû '11

25 Aoû '11

Hi, Today (in one hour) we have Steven Pigeon giving us a brief primer on compression methods. Be there. Time/Date: 14h00 August 25th Location: LISA Lab (AA3256) Title: Compression 101: A brief Introduction The goal of the talk is to introduce the goals of data compression and its basic concepts. We will illustrate the concepts using the different stages of a typical codec (for CODer/DECorder pair) such as transformation, analysis, quantization and entropy coding. (Snacks provided by Steven, yay! :-)) Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Tomorrow!
by Aaron Courville 24 Aoû '11

24 Aoû '11

Hi, Tomorrow we will have Steven Pigeon give us a brief primer on compression methods. I hope to see you there. Time/Date: 14h00 August 25th Location: LISA Lab (AA3256) Title: Compression 101: A brief Introduction The goal of the talk is to introduce the goals of data compression and its basic concepts. We will illustrate the concepts using the different stages of a typical codec (for CODer/DECorder pair) such as transformation, analysis, quantization and entropy coding. (Snacks provided by Steven, yay! :-)) Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Thursday!
by Aaron Courville 16 Aoû '11

16 Aoû '11

Hi Gang, This week we will return to our regularly schedule tea talks with a talk by David Rim. Title and abstract are below. Hope to see you there. Time: 14h00, Thursday August 18th Location: LISA lab (AA3256) Title: Learning Identity-Invariant Features for Expression Recognition with 3D CGI Models for Semi-Supervised Learning Abstract: I plan to propose a method for learning features for expression recognition by leveraging work done in the graphics community for modeling and generating realistic human faces and related motion-capture technology. I will demonstrate that these models provide expression representations which are far more compact and more importantly, easier to interpret and code than Facial Action Coding System Action Units. I will demonstrate a labeling process which I believe will lead to more accurate labels, and show initial experiments using simple linear models for feature learning. I will propose a convolutional network model to learn better features (filter maps) across models for different "identities" as well as using unlabeled human faces. Finally, I will propose a semi-supervised model which can use these features to learn an inductive function which maps 2D facial images to the expression representation using these learned features inductively. I hope a useful discussion will follow. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk this week
by Aaron Courville 26 Jul '11

26 Jul '11

Hi All, This week we will have Navdeep Jaitly, a student with Geoffrey Hinton, who will tell us about his work on Transforming Autoencoders. See you there. Where: LISA lab (AA3256) When: 14h00, Thursday, July 28th 2011 Title: Transforming Autoencoders for speech. Speech recognition is dominated by use of Gaussian Mixture Model / Hidden Markov Models trained on mel frequency cepstral representations of speech. Psychologists have long argued that highy-quality recognition would be facilitated by finding acoustic events or landmarks that have well-defined onset times, amplitudes and rates in addition to being present or absent. We introduce a new way of learning such acoustic events by using a new type of autoencoder that is given both a spectrogram and desired global transformation and learns to output the transformed spectrogram. By specifying the global transformation in the appropriate way, we can force the autoencoder to extract accoustic events that, in addition to a probability of being present, have explicit onset times, amplitudes and rates. This makes it much easier to compute relationships between acoustic events. I'll talk about the work we have done and the directions we are currently pursuing. This is joint work with Geoffrey Hinton. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk this Thursday
by Aaron Courville 19 Jul '11

19 Jul '11

Hi Gang, This Thursday we will have Tomáš Mikolov talk about his recent work with recurrent neural networks. Hope to see you all there. Time & Date: 14h00 Thursday July 21st. Location: LISA Lab (AA3256) Title: Recent advances in neural network based language modeling Abstract: I plan to describe novel techniques for training neural network language models. I will introduce the recurrent architecture, class based neural net model, and simple maximum entropy model that can be trained jointly with the neural net model. Together, these techniques allow training of very good language models on hundreds of millions of words. Results after application to the state of the art speech recognition systems will be presented, as well as comparison to other popular language modeling techniques. Next, I'd like to present some ideas how to further extend and improve the existing models, and how to possibly make them closer to the human performance. I hope for some discussion about this part. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Tomorrow!
by Aaron Courville 13 Jul '11

13 Jul '11

Hi All, This week we will have a Tea Talk by Razvan describing some of his recent modifications of Theano. Hope to see you there. Time / Date: 14h00 Thursday July 14th (Tomorrow) Location: LISA LAB (AA3256) Title: Better graph visualization and while loops in Theano Abstract: In the same spirit as Philippe's presentation, I want to go over a few new features that I have currently in my branch of Theano and that I hope to push to the trunk soon. My intent is to have a short demo, talk a bit about what works, what still doesn't, how to use them what is left to be done. I intent to finish my short presentation with a open discussion on Theano if participants are up to it. I'm interested to get some feedback on what I've been working on, and get a general sense of what people in the lab actually think about Theano, what are things people would want to see done in the near future, etc. Hope to see you there. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Thursday!
by Aaron Courville 08 Jul '11

08 Jul '11

Hi All, This week we will have Philippe Hamel present a python library that he and Simon have been working on. Hope to see you all there. Where: LISA lab (AA3256) When: 14h00, Thursday, July 7th, 2011 Title/Abstract: Developing tools for hyper-parameter visualization As the complexity of machine learning algorithms becomes greater, the number of hyper-parameters tends to do the same. Optimizing such algorithms often involves tedious exploration of a high dimensional hyper-parameter space. Typically, results from experiments are stored in a SQL database and results can be analyzed by looking at sql queries in a table. With this kind of view, it can be hard to see correlations between parameters, or to understand trends in the data. Using plots instead of tables can allow for better understanding of the hyper-parameter space, but might require more coding. So, a tool that allows us to easily plot and visualize the hyper-parameter space could help to guide our experiments in the right direction. In this tea talk, I would like to present a small python library that Simon and I have been working on and using to visualize our experiments. It is basically a bridge between a jobman SQL database and pylab. It allows dynamic filtering of the data, applying functions to transform and/or combine columns of data, and more! We have a few ideas on features that could be added, and we would like to get your input. We would like to give you a quick demo of what the library can do for now. We'd love some feedback from you. We are interested to know if people from LISA might be interested to use this tool, or even contribute to it. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

2 1

Correction: Tea Talk Tomorrow (Friday)
by Aaron Courville 16 Jui '11

16 Jui '11

[Corrected title and abstract below] Hi Gang, Tomorrow we will have a tea talk by David Rim (PhD student of Chris Pal) telling us about some recent vision work. Hope to see you all there. Where: LISA lab (AA3256) When: Friday June 17th, 14h00 Title: Realtime Performance-Based Facial Animation Thibaut Weise, Sofien Bouaziz, Hao Li, Mark Pauly EPL ACM Transactions on Graphics, Proceedings of the 38th ACM SIGGRAPH Conference and Exhibition 2011, 08/2011 – SIGGRAPH 2011 http://lgg.epfl.ch/~weise/publications/papers/rpbfa_siggraph2011_small.pdf Abstract: This paper presents a system for performance-based character ani- mation that enables any user to control the facial expressions of a digital avatar in realtime. The user is recorded in a natural envi- ronment using a non-intrusive, commercially available 3D sensor. The simplicity of this acquisition device comes at the cost of high noise levels in the acquired data. To effectively map low-quality 2D images and 3D depth maps to realistic facial expressions, we intro- duce a novel face tracking algorithm that combines geometry and texture registration with pre-recorded animation priors in a single optimization. Formulated as a maximum a posteriori estimation in a reduced parameter space, our method implicitly exploits temporal coherence to stabilize the tracking. We demonstrate that compelling 3D facial dynamics can be reconstructed in realtime without the use of face markers, intrusive lighting, or complex scanning hardware. This makes our system easy to deploy and facilitates a range of new applications, e.g. in digital gameplay or social interactions. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk Tomorrow (Friday)
by Aaron Courville 16 Jui '11

16 Jui '11

Hi Gang, Tomorrow we will have a tea talk by David Rim (PhD student of Chris Pal) telling us about some recent vision work. Hope to see you all there. Where: LISA lab (AA3256) When: Friday June 17th, 14h00 Title / Abstract: Accurate and robust registration for in-hand modeling Weise, T. and Leibe, B. and Van Gool, L. Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on, 2008 http://www.vision.ee.ethz.ch/~bleibe/papers/weise-inhandscanning-cvpr08.pdf Abstract This paper presents a system for performance-based character animation that enables any user to control the facial expressions of a digital avatar in realtime. The user is recorded in a natural environment using a non-intrusive, commercially available 3D sensor. The simplicity of this acquisition device comes at the cost of high noise levels in the acquired data. To effectively map low-quality 2D images and 3D depth maps to realistic facial expressions, we introduce a novel face tracking algorithm that combines geometry and texture registration with pre-recorded animation priors in a single optimization. Formulated as a maximum a posteriori estimation in a reduced parameter space, our method implicitly exploits temporal coherence to stabilize the tracking. We demonstrate that compelling 3D facial dynamics can be reconstructed in realtime without the use of face markers, intrusive lighting, or complex scanning hardware. This makes our system easy to deploy and facilitates a range of new applications, e.g. in digital gameplay or social interactions. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Reminder: Tea Talk (NOTE NEW LOCATION)
by Aaron Courville 08 Jui '11

08 Jui '11

Hi Gang, This week we welcome the return of the tea talk with Frédéric Bastien giving a practice run of his Theano tutorial. Location: AA3195 Time: 14h00 Thursday June 9th GPU Programming made Easy Instructor: Frédéric Bastien, Université de Montréal Abstract: In this tutorial, the student will learn to program GPUs using tools such as PyCUDA and Theano to access the power of GPU computing without the complexity inherent to CUDA programming. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk
by Aaron Courville 07 Jui '11

07 Jui '11

Hi Gang, This week we welcome the return of the tea talk with Frédéric Bastien giving a practice run of his Theano tutorial. Location: TBA Time: 14h00 Thursday June 9th GPU Programming made Easy Instructor: Frédéric Bastien, Université de Montréal Abstract: In this tutorial, the student will learn to program GPUs using tools such as PyCUDA and Theano to access the power of GPU computing without the complexity inherent to CUDA programming. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Tea Talk this week! (Note new time)
by Aaron Courville 10 Mai '11

10 Mai '11

Hi All, This week Yoshua will share some of his new ideas based on the recent work of Hinton et al. Hope to see you there! Time/Date: 14h00 Thursday May 12th Location: LISA Lab (AA3256) Abstract: I will talk about the transformer networks recently proposed by Geoff in this attached to-appear ICANN paper, and ideas that I've had to move even further and do things that could be very interesting to Ubisoft. What can I say more than what I had in my NIPS workshops video: it is how the brain works! so I encourage all to attend and join a big brainstorming session. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0