Lisa_seminaires Mars 2017

lisa_seminaires@iro.umontreal.ca

8 participants
8 discussions

Brian Ziebart on "Supervised Machine Learning as an Adversarial Game", March 31, 13:45, AA6214
by Junyoung Chung 31 Mar '17

31 Mar '17

Hi all, the next tea-talk will be given by Prof. Brian Ziebart from the University of Illinois at Chicago. The talk will be held at 13:45 & AA6214. If you would like to chat with Brian, please fill your names and available time-slots on the below spreadsheet: https://docs.google.com/spreadsheets/d/14PpRI2WYl_zPlLPGC5SPQu4fFbJDFHIIXt4… Title: Supervised Machine Learning as an Adversarial Game Abstract: A standard approach to supervised machine learning is to choose the form of a predictor and to then optimize its parameters based on training data. Approximations of the predictor's performance measure are often required to make this optimization problem tractable. Instead of approximating the performance measure and using the exact training data, this talk explores adversarial approximations of the training data while optimizing the exact performance measure. This formulation ensures stronger theoretical guarantees (e.g., Fisher consistency) than other surrogate loss methods, better aligns predictor construction with univariate and multivariate performance measures found in computer vision, information retrieval, and rank learning tasks, and provides flexibility for addressing sample selection bias for improved active learning. Bio: Brian Ziebart is an Assistant Professor in the Department of Computer Science at the University of Illinois at Chicago. He earned his PhD in Machine Learning from Carnegie Mellon University where he was also a postdoctoral fellow. His interests lie in the intersections between machine learning, game theory, and decision theory. He has published over 25 articles in leading machine learning and artificial intelligence venues, including one that received a Best Paper award at the International Conference on Machine Learning. Best, --Junyoung

3 6

Talk CHANGE of LOCATION Today 2:30pm - 3:30pm
by Aaron Courville 28 Mar '17

28 Mar '17

This is a reminder that this afternoon we will have a talk by Lluís Castrejón, a visiting student from Toronto currently working with Raquel Urtasun and Sanja Fidler. He will be presenting his recently accepted CVPR (oral) paper. Time: 2:30pm - 3:30pm. Location: AA5340 (NOTE NEW LOCATION). Title: Annotating Object Instances with a Polygon-RNN Abstract: We propose an approach for semi-automatic annotation of object instances. While most current methods treat object segmentation as a pixel-labeling problem, we here cast it as a polygon prediction task, mimicking how most current datasets have been annotated. In particular, our approach takes as input an image crop and produces a vertex of the polygon, one at a time, allowing a human annotator to interfere at any time and correct the point. Our model easily integrates any correction,producing as accurate segmentations as desired by the annotator. We show that our annotation method speeds up the annotation process by factor of 4.7 across all classes in Cityscapes, while achieving 78.4% agreement in IoU with original ground-truth, matching the typical agreement between human annotators. For cars, our speed-up factor is even higher, at 7.3 for an agreement of 82.2%. We further show generalization capabilities of our approach on unseen datasets. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Talk (Special Time) Tuesday 2:30pm - 3:30pm
by Aaron Courville 27 Mar '17

27 Mar '17

On Tuesday (tomorrow) we will have a talk by Lluís Castrejón, a visiting student from Toronto currently working with Raquel Urtasun and Sanja Fidler. He will be presenting his recently accepted CVPR (oral) paper. If you would like to meet with him please let me know. Time: 2:30pm - 3:30pm. Location: AA6214. Title: Annotating Object Instances with a Polygon-RNN Abstract: We propose an approach for semi-automatic annotation of object instances. While most current methods treat object segmentation as a pixel-labeling problem, we here cast it as a polygon prediction task, mimicking how most current datasets have been annotated. In particular, our approach takes as input an image crop and produces a vertex of the polygon, one at a time, allowing a human annotator to interfere at any time and correct the point. Our model easily integrates any correction,producing as accurate segmentations as desired by the annotator. We show that our annotation method speeds up the annotation process by factor of 4.7 across all classes in Cityscapes, while achieving 78.4% agreement in IoU with original ground-truth, matching the typical agreement between human annotators. For cars, our speed-up factor is even higher, at 7.3 for an agreement of 82.2%. We further show generalization capabilities of our approach on unseen datasets. Cheers, Aaron -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

[Tea-Talk] Andreas Moshovos, March 24, 13:45, AA6214
by Dzmitry Bahdanau 24 Mar '17

24 Mar '17

Hi all, Our next tea-talk will be given by Dr. Andreas Moshovos *on March 24, at 13:45 (new time!), room AA6214*, who is a professor at University of Toronto. Hope you see many of you there! *Please let me know if you would like to speak to Andreas after the talk. *You can directly put your name on the spreadsheet below: https://docs.google.com/spreadsheets/d/1vVn0eOAlTmkmd-DpEwuMD2NEQH_H8NLlY5Y… *Title: *Exploiting Value Content to Accelerate Inference with Convolutional Neural Networks *Abstract: *Sufficiently capable computing hardware is essential for practical applications of Deep Learning. Until very recently computing hardware capabilities have been increasing at an exponential rate. As a result, around 2010 computing hardware capabilities reached the level necessary to demonstrate Deep Learning’s true potential. Unfortunately, semiconductor technology scaling, the keep enabler of this past exponential growth in capabilities, has slowed down dramatically. Fortunately, specialized computing hardware design has the potential to deliver another 2 to 3 orders of improvements in computing capabilities. Our goal is to develop the techniques necessary for boosting computing hardware capabilities thus enabling further innovation in Deep Learning. We are developing specialized computing hardware for Deep Learning Networks whose key feature is that they are value-based. We have been developing value-based accelerators that take advantage of expected properties in the runtime calculated value stream of Deep Learning Networks such as the value distribution of activations, or even their bit content. Using image classification convolutional neural networks, we have demonstrated 2 to 3 orders of magnitude execution time improvements over conventional graphics processor hardware and up to 4.5x improvements over a state-of-the-art accelerator. In this talk we will review the need for specialized computing hardware for Deep Learning and summarize our efforts. We will also briefly touch upon the recently approved NSERC COHESA Strategic Partnership Network on Hardware Acceleration for Machine Learning. NSERC COHESA brings together 19 Researchers across multiple Canadian Universities and 8 Industrial Partners. *Bio: *Andreas Moshovos teaches how to design and optimize computing hardware engines at the University of Toronto where he has the privilege of collaborating with several talented students on techniques to improve execution time, energy efficiency and cost for computing hardware. He has also taught at Northwestern University, USA, the University of Athens, Greece, the Hellenic Open University, Greece, and as an invited professor at the École Polytechnique Fédérale de Lausanne, Switzerland. He has received the ACM SIGARCH Maurice Wilkes award in 2010, an NSF CAREER Award in 2000, two IBM Faculty awards, a Semiconductor Research Corporation Inventor recognition award, and a MICRO Hall of Fame award. He has served at the Program Chair for the ACM/IEEE International Symposium on Microarchiteture and the IEEE International Symposium on the Performance Analysis of Systems and Software. He studied computer science at the University of Crete, Greece, at New York University, USA, and at the University of Wisconsin-Madison, USA. Best, Dima

1 2

[Tea-Talk] Devon Hjelm & Laurent Dinh, March 17, 14:30, AA6214
by Dzmitry Bahdanau 17 Mar '17

17 Mar '17

Hi all, At the next tea-talk Devon Hjelm (post-doc) and Laurent Dinh (PhD student) from MILA will present their ICML submissions. Please come in great numbers on *March 17 to AA6214 at 13:30*! The details are below. *Speaker:* Devon Hjelm *Title:* Boundary-Seeking Generative Adversarial Networks *Abstract: *We introduce a novel approach to training generative adversarial networks (GANs, Goodfellow et al., 2014), which stems from reinterpreting the generator objective to match a target distribution that converges to the data distribution at the limit of a perfect discriminator. This objective can be interpreted as training the generator to produce samples that lie on the decision boundary of the current discriminator in training, and we call this method boundary-seeking GANs (BS-GAN). This approach can be used to train a generator with discrete output in the case that the generator is parametrized by a conditional distribution, and we demonstrate this with discrete image data. We also observe that the Gumbel-softmax trick does not work for training GANs with discrete data. Finally, our approach suggests a new objective function even for continuously valued data, and we demonstrate this with common image datasets. *Speaker:* Laurent Dinh *Title: *Sharp Minima Can Generative For Deep Nets *Abstract: *Despite their overwhelming capacity to overfit, deep learning architectures tend to generalize relatively well to unseen data, allowing them to be deployed in practice. However, explaining why this is the case is still an open area of research. One standing hypothesis that is gaining popularity, e.g. Hochreiter & Schmidhuber (1997); Keskar et al. (2017), is that the flatness of minima of the loss function found by stochastic gradient based methods results in good generalization. This paper argues that most notions of flatness are problematic for deep models and can not be directly applied to explain generalization. Specifically, when focusing on deep networks with rectifier units, we can exploit the particular geometry of parameter space induced by the inherent symmetries that these architectures exhibit to build equivalent models corresponding to arbitrarily sharper minima. Or, depending on the definition of flatness, it is the same for any given minimum. Furthermore, if we allow to reparametrize a function, the geometry of its parameters can change drastically without affecting its generalization properties. Best, Dima

2 3

Jörn Diedrichsen on "The brain’s GPU? In search of the cerebellum’s universal computation", March 10, 13:30, AA6214
by Junyoung Chung 10 Mar '17

10 Mar '17

Hi all, the next tea talk will be given by a guest speaker, Jörn Diedrichsen from UWO. Same place, same time: March 10, at 13:30-14:30, room AA6214. Hope to see many of you there! Title: The brain’s GPU? In search of the cerebellum’s universal computation Abstract: The cerebellum exhibits a highly specialized and uniform neuronal circuitry, which likely evolved to solve very specific problems in sensory-motor control. In the human brain, the cerebellar circuitry has dramatically expanded, and it contributes here to virtually every possible cognitive function. But what is this elusive computation that the cerebellum contributes to cortical processing? I will show some insights from our research on motor control and learning, which shows that the cerebellum is critically involved in prediction and fast error-based learning. I will then talk about the first steps that we have recently taken to test computational theories about the function of the cerebellum across different cognitive domains using functional imaging. Bio: Jörn Diedrichsen is a Wester Research Chair in the Brain and Mind Institute at Western University, and Professor of Statistics and Computer Science. His laboratory studies the cerebellar function and motor control and motor learning in the human brain. He has also developed a number of novel multivariate analysis techniques for functional imaging data. Best, --Junyoung

3 4

[Extra Tea-Talk] Martin Arjovsky, March 7, 14:30, AA6214
by Dzmitry Bahdanau 07 Mar '17

07 Mar '17

Hi all, This Tuesday (March 7) we will have an extra tea-talk by *Martin Arjovsky*. Please note that the time is different (14:30), but the place is the same (AA6214). Hope to see many of you there! *Title: *On Different Distances Between Distributions and Generative Adversarial Networks *Abstract: *Generative adversarial networks (GANs) are notoriously difficult to train. At the core of it, we show that these problems arise naturally when trying to learn distributions whose support lie in low dimensional manifolds. We show how these problems are consequences of trying to optimize the classical divergences (KL, JSD, etc) between our real and data distribution, and that these are symptoms of a more general phenomenon, pointing towards the inefficacy of the usual divergences in certain settings. After that, we bring into play the Wasserstein distance, which we prove doesn't suffer from the same behaviour, and provide a first step on an algorithm that tries to approximately optimize this distance. Dima

2 2

[TeaTalk] Zhouhan Lin, March 3, 13:30, AA6214
by Dzmitry Bahdanau 03 Mar '17

03 Mar '17

Hi all, Our next speaker is Zhouhan Lin, a PhD student from MILA. The time and place are usual: March 3, at 1:30pm, room AA6214. Home to see many of you there! *Title: *A Structured Self-Attentive Sentence Embedding *Abstract: *This paper proposes a new model for extracting an interpretable sentence embedding by introducing self-attention. Instead of using a vector, we use a 2-D matrix to represent the embedding, with each row of the matrix attending on a different part of the sentence. We also propose a self-attention mechanism and a special regularization term for the model. As a side effect, the embedding comes with an easy way of visualizing what specific parts of the sentence are encoded into the embedding. We evaluate our model on 3 different tasks: author profiling, sentiment classification and textual entailment. Results show that our model yields a significant performance gain compared to other sentence embedding methods in all of the 3 tasks Dima

1 3

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

Lisa_seminaires Mars 2017