Lisa_seminaires Mai 2017

lisa_seminaires@iro.umontreal.ca

4 participants
8 discussions

Christian Gagné and Jean-François Lalonde from U. of Laval.
by Junyoung Chung 02 Jui '17

02 Jui '17

Hi all, we have two tea-talks (each is a 30-min talk) this week! Both speakers are from Univ. of Laval. You will find more information below. Talk by Christian Gagné Time: 13:45~14:15 Title: Alternating Direction Method of Multipliers for Sparse Convolutional Neural Networks Abstract: The storage and computation requirements of Convolutional Neural Networks (CNNs) can be prohibitive for exploiting these models over low-power or embedded devices. We are proposing to reduce the computational complexity of the CNNs by minimizing an objective function, including the recognition loss that is augmented with a sparsity-promoting penalty term. The sparsity structure of the network is identified using the Alternating Direction Method of Multipliers (ADMM), which is widely used in large optimization problems. This method alternates between promoting the sparsity of the network and optimizing the recognition performance, which allows us to exploit the two-part structure of the corresponding objective functions. In particular, we take advantage of the separability of the sparsity-inducing penalty functions to decompose the minimization problem into sub-problems that can be solved sequentially. Applying our method to a variety of state-of-the-art CNN models, our proposed method is able to simplify the original model, generating models with less computation and fewer parameters, while maintaining and often improving generalization performance. Accomplishments on a variety of models strongly verify that our proposed ADMM-based method can be a very useful tool for simplifying and improving deep CNNs. Short bio: Christian Gagné is professor at the Electrical Engineering and Computer Engineering Department of Université Laval since 2008. He is member of the Computer Vision and Systems Laboratory, the REPARTI FRQNT Strategic Cluster, and the Big Data Research Centre. He completed a Ph.D. in Electrical Engineering (Université Laval) in 2005 before being postdoctoral fellow at the INRIA Saclay - Île-de-France (France) and the University of Lausanne (Switzerland) in 2005-2006. He was research associate for Informatique WGZ Inc. (2006-2007) and MacDonald, Dettwiler, and Associates Ltd. (2007-2008). His research interests are on the engineering of intelligent systems, especially systems based on machine learning, evolutionary computing, and their application to pattern recognition, smart sensors, smart grids, and simulation of public health processes. Talk by Jean-François Lalonde Time: 14:15~14:45 Title: Learning to Predict Illumination from a Single Image Abstract: We propose an automatic method to infer high dynamic range illumination from a single, limited field-of-view, low dynamic range photograph of a scene. In contrast to previous work that relies on specialized image capture, user input, and/or simple scene models, we train an end-to-end deep neural network that directly regresses a limited field-of-view photo to HDR illumination, without strong assumptions on scene geometry, material properties, or lighting. This allows us to automatically recover high-quality HDR illumination estimates that significantly outperform previous state-of-the-art methods. Consequently, using our illumination estimates for applications like 3D object insertion, we can achieve results that are photo-realistic. Bio: Jean-François Lalonde is an Assistant Professor in Electrical and Computer Engineering at Laval University, Quebec City, since 2013. Previously, he was a Post-Doctoral Associate at Disney Research, Pittsburgh. He received a B.Eng. degree in Computer Engineering with honors from Laval University, Canada, in 2004. He earned his M.S at the Robotics Institute at Carnegie Mellon University in 2006 under Prof. Martial Hebert and received his Ph.D., also from Carnegie Mellon, in 2011 under the supervision of Profs. Alexei A. Efros and Srinivasa G. Narasimhan. His thesis, titled "Understanding and Recreating Appearance under Natural Illumination," won the 2010-11 CMU School of Computer Science Distinguished Dissertation Award. After graduation, he became a Computer Vision Scientist at Tandent, Inc., where he helped develop LightBrush™, the first commercial intrinsic imaging application. He also introduced intrinsic videos at SIGGRAPH 2012 while at Tandent. His research focuses on lighting-aware image understanding and synthesis by leveraging large amounts of data. More at www.jflalonde.ca. Best, --Junyoung

1 3

Reminder: Talk Today at 2pm in AA3195.
by Aaron Courville 31 Mai '17

31 Mai '17

We will have a talk today at 2pm. Our speaker is Michael Chang from MIT, working with Josh Tenenbaum. Date / Time: Wednesday May 31st, 2pm-3pm Location: AA3195 Title: *Learning Visual and Physical Models of the Environment* Abstract: An intelligent agent can leverage a model of the environment as a prior to accelerate future learning and generalize beyond its own experience. I view building such models as learning simulator programs that infer latent properties and generate predictions from observation. How can the agent learn the primitives for these programs from observation? What are the means of combination that operate on these primitives that allow zero-shot generalization to come naturally? In this talk, I discuss two of my recent work that propose potential directions for tackling these questions in the context of learning visual concepts and intuitive physics. First, I present an algorithm for learning factorized symbolic representations from raw visual data. I show that these representations capture latent factors of variation that can be manipulated like a graphics code. Next, I present a framework for learning predictive models of intuitive physics. With two-dimensional worlds of bouncing balls, I demonstrate that this framework generalizes to variable object count and variable scene configurations with only spatially and temporally local computation. I conclude by describing open research questions motivated by the results from these approaches. Bio: Michael Chang is a senior in Computer Science at MIT, researching in Professor Joshua Tenenbaum's Computational Cognitive Science Group. He is interested in building algorithms for learning compositional programs and bridging the strengths of symbolic and neural representations. At MIT, he has worked on unsupervised learning of symbolic visual concepts and of physical dynamics. Michael has spent time at the University of Michigan researching with Professor Honglak Lee. Michael will be pursuing a Ph.D. in Computer Science at U.C. Berkeley beginning fall 2017. Links to papers and code are here: http://mbchang.github.io. -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email: Aaron.Courville(a)gmail.com

1 0

Talk Tomorrow!!!
by Aaron Courville 30 Mai '17

30 Mai '17

Everyone, We will have a talk tomorrow at 2pm (note unusual day/time). Our speaker is Michael Chang from MIT, working with Josh Tenenbaum. Date / Time: Wednesday May 31st, 2pm-3pm Location: AA3195 Title: *Learning Visual and Physical Models of the Environment* Abstract: An intelligent agent can leverage a model of the environment as a prior to accelerate future learning and generalize beyond its own experience. I view building such models as learning simulator programs that infer latent properties and generate predictions from observation. How can the agent learn the primitives for these programs from observation? What are the means of combination that operate on these primitives that allow zero-shot generalization to come naturally? In this talk, I discuss two of my recent work that propose potential directions for tackling these questions in the context of learning visual concepts and intuitive physics. First, I present an algorithm for learning factorized symbolic representations from raw visual data. I show that these representations capture latent factors of variation that can be manipulated like a graphics code. Next, I present a framework for learning predictive models of intuitive physics. With two-dimensional worlds of bouncing balls, I demonstrate that this framework generalizes to variable object count and variable scene configurations with only spatially and temporally local computation. I conclude by describing open research questions motivated by the results from these approaches. Bio: Michael Chang is a senior in Computer Science at MIT, researching in Professor Joshua Tenenbaum's Computational Cognitive Science Group. He is interested in building algorithms for learning compositional programs and bridging the strengths of symbolic and neural representations. At MIT, he has worked on unsupervised learning of symbolic visual concepts and of physical dynamics. Michael has spent time at the University of Michigan researching with Professor Honglak Lee. Michael will be pursuing a Ph.D. in Computer Science at U.C. Berkeley beginning fall 2017. Links to papers and code are here: http://mbchang.github.io. -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email: Aaron.Courville(a)gmail.com

1 0

[Tea Talk] Marco Pedersoli, May 26, AA5340, 13:45
by Dzmitry Bahdanau 26 Mai '17

26 Mai '17

Hi all, Our next speaker is Marco Pedersoli who is a professor at the École de technologie supérieure of Montreal (ETS). Hope to see many of you there! *Title:* Areas of Attention for Image Captioning *Abstract: *We propose “Areas of Attention”, a novel attention-based model for automatic image captioning. Our approachmodels the dependencies between image regions, caption words, and the state of an RNN language model, using three pairwise interactions. In contrast to previous attention-based approaches that associate image regions only to the RNN state, our method allows a direct association between caption words and image regions. During training these associations are inferred from image-level captions, akin to weakly-supervised object detector training. These associ- ations help to improve captioning by localizing the corresponding regions during testing. We also propose and com- pare different ways of generating attention areas: CNN activation grids, object proposals, and spatial transformers nets applied in a convolutional fashion. Spatial transformers give the best results. They allow for image specific at- tention areas, and can be trained jointly with the rest of the network. Our attention mechanism and spatial transformer attention areas together yield state-of-the-art results on the MSCOCO dataset. *Bio:* Marco Pedersoli obtained his Ph.D. from the Autonomous University Of Barcelona (june 2012, with distinction and best thesis award) on Hierarchical Multi-resolution Detection of objects in images. He completed a postdoctoral fellowship at the KU Leuven, where he has developed several innovative approaches for object detection, action classification and pose estimation based on weakly-supervised methods. In September 2015 he moved to INRIA-Grenoble, where he has developed new techniques for the automatic description and understanding of images. From February 2017 he is assistant professor at the École de technologie supérieure of Montreal. He has published more than 20 articles in peer-reviewed internationals journals and top peer-reviewed conferences in computer vision. Dima

1 1

No tea-talk this week
by Junyoung Chung 09 Mai '17

09 Mai '17

Hi all, we do not have a talk this week due to NIPS submission deadline. See you after the deadline and wish you a good luck! Best, --Junyoung

1 0

Fwd: Yaroslav Ganin, PREDOC, 8 mai.
by Yoshua Bengio 08 Mai '17

08 Mai '17

Yaroslav Ganin will present his thesis proposal (oral predoctoral exam) this afternoon at 3pm at rm 3195. Please come! ---------- Message transféré ---------- De : <tappa(a)iro.umontreal.ca> Date : 3 mai 2017 à 12:09 Objet : Yaroslav Ganin, PREDOC, 8 mai. À : seminaires(a)iro.umontreal.ca Cc : ganiniar(a)iro.umontreal.ca Bonjour Vous êtes tous cordialement invités à assister à la présentation du sujet de recherche de Yaroslav Ganin (partie orale de l'examen prédoctoral). Title : Natural image processing and synthesis using deep neural networks Jury : Aaron Courville, Yoshua Bengio et Alain Tapp Lieu : AA 3195 Date: lundi le 8 mai à 15h00. -- Alain Tapp Professeur titulaire DIRO, Université de Montréal http://www.iro.umontreal.ca/~tappa/

1 0

Timnit Gebru (Stanford CV Lab, PhD candidate), May 5, 13:45, AA5340
by Junyoung Chung 05 Mai '17

05 Mai '17

Hi all, the next tea-talk will be given by Timnit Gebru from Stanford university. The room will be AA5340, and the talk will start at 13:45. Here is the detail of the talk. *Title: *Using Deep Learning and Google Street View to Estimate the Demographic Makeup of the US *Abstract: *Targeted socio-economic policies require an accurate understanding of a country’s demographic makeup. To that end, the United States spends more than 1 billion dollars a year gathering census data such as race, gender, education, occupation and unemployment rates. Compared to the traditional method of collecting surveys across many years which is costly and labor intensive, data-driven, machine learning driven approaches are cheaper and faster—with the potential ability to detect trends in close to real time. In this work, we leverage the ubiquity of Google Street View images and develop a computer vision pipeline to predict income, per capita carbon emission, crime rates and other city attributes from a single source of publicly available visual data. We first detect cars in 50 million images across 200 of the largest US cities and train a model to determine demographic attributes using the detected cars. To facilitate our work, we used a graph-based algorithm to collect the largest and most challenging fine-grained dataset reported to date consisting of over 2600 classes of cars comprised of images from Google Street View and other web sources. Our prediction results correlate well with ground truth income (r=0.82), race, education, voting, sources investigating crime rates, income segregation, per capita carbon emission, and other market research. Finally, we learn interesting relationships between cars and neighborhoods allowing us to perform the first large-scale sociological analysis of cities using computer vision techniques. *Bio: *I am a PhD student in the Stanford Artificial Intelligence Laboratory, studying computer vision under Fei-Fei Li. My main research interest lies in data mining large-scale publicly available images to gain sociological insight, and working on computer vision problems that arise as a result. Some of these include fine-grained image recognition, scalable annotation of images, and domain adaptation. Prior to joining Fei-Fei's lab I worked at Apple designing circuits and signal processing algorithms for various Apple products including the first iPad. I also spent an obligatory year as an entrepreneur (as all Stanford undergrads seem to do). My research is supported by the NSF foundation GRFP fellowship and currently the Stanford DARE fellowship. Best, --Junyoung

1 3

Fwd: présentation prédoc III - Yifan Nie
by Yoshua Bengio 03 Mai '17

03 Mai '17

This should be of interest to many of us, from our colleagues in the DIRO's NLP group. ---------- Forwarded message ---------- From: nie <nie(a)iro.umontreal.ca> Date: 2017-05-03 17:24 GMT-04:00 Subject: présentation prédoc III - Yifan Nie To: seminaires(a)iro.umontreal.ca Vous êtes cordialement invités à cette présentation prédoc III, par: Yifan Nie (RALI) Tailored Deep Learning Techniques for Information Retrieval Vendredi 5 mai 2017, 11:00, salle Z-200, pavillon Claire McNicoll Traditional information retrieval systems employing bag-of-words representations often lead to poor results. More recently, learning-to-rank (L2R) approaches have successfully incorporated the idea of learning a ranking function. However, the features used to represent documents and queries in L2R are still human-crafted. In this research, we will incorporate representation learning developed in the area of deep learning. The proposed framework is tailored to IR tasks, which combines representation learning and ranking-function learning. To implement the framework, several key problems need to be addressed: 1. Query-dependent representation learning; 2. Using multiple representations in a deep net; 3. Generating pseudo training data for IR. In this presentation, we will propose approaches to these problems and report some preliminary results. Jury: Aaron Courville Philippe Langlais Jian-Yun Nie (directeur)

1 0

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

Lisa_seminaires Mai 2017