Lisa_seminaires

lisa_seminaires@iro.umontreal.ca

396 discussions

[Tea Talk] Viral Shah (Julia Computing) Fri September 14 2018 10:30 Z315
by Michael Noukhovitch 10 Sep '18

10 Sep '18

This week we have *Viral Shah * from * Julia Computing * giving a talk on *Fri September 14 2018* at *10:30* in room *Z315 (in Pavilion Claire McNicoll)*. It's not our usual room but it'll be a regular room for the next two months :) Will this talk be streamed <https://mila.bluejeans.com/809027115/webrtc>? Yes And you can sign up to meet the speaker here: if there's a tea talk, you should always GOTO it Michael *TITLE* On Machine Learning and Programming Languages *KEYWORDS *ML/DL infrastructure, Programming Languages *ABSTRACT* We ask, what might the ideal ML language of the future look like? Our thoughts are published in this blog post: https://julialang.org/blog/2017/12/ml&pl As programming languages (PL) people, we have watched with great interest as machine learning (ML) has exploded -- and with it, the complexity of ML models and the frameworks people are using to build them. State-of-the-art models are increasingly programs, with support for programming constructs like loops and recursion, and this brings out many interesting issues in the tools we use to create them -- that is, programming languages. While machine learning does not yet have a dedicated language, several efforts are effectively creating hidden new languages underneath a Python API (like TensorFlow) while others are reusing Python as a modeling language (like PyTorch). We'd like to ask -- are new ML-tailored languages required, and if so, why? Now that Julia 1.0 is released, we will also discuss how Julia evolved to get where it is today, and how it might evolve to taking on some of the challenges posed by machine learning *BIO* Dr. Viral Shah is a co-creator of the Julia project and Co-founder and CEO of Julia Computing. He has had a long-term track record of building open-source software. Apart from Julia, he is also co-creator of Circuitscape, an open-source program which borrows algorithms from electronic circuit theory for ecological conservation. In the Government of India, he was an early member of the country’s national ID project - Aadhaar, where his work on re-architecting India’s social security systems led to a significant increase in social and financial inclusion, while simultaneously saving the exchequer over a billion dollars in slippage. The experiences of implementing technology at such scale for a billion people are collected in his book: Rebooting India. Viral has a Ph. D. from the University of California at Santa Barbara, in Computer Science.

1 1

[Tea Talk] Geoff Gordon (MSR Montreal) Fri September 7 2018 10:30 AA3195
by Michael Noukhovitch 07 Sep '18

07 Sep '18

This week we are super lucky to have *Geoff Gordon*, Research Director of * MSR Montreal * giving a talk on *Fri September 7 2018* at *10:30* in room *AA3195* Will this talk be streamed <https://mila.bluejeans.com/809027115/webrtc>? *No* Geoff will be available to meet in the afternoon! Sign up here <https://calendar.google.com/calendar/selfsched?sstoken=UVBVTVF0X25Nd09LfGRl…> Knowing that this talk is likely awesome and that you probably come to talks given they're awesome, you're probably going to the talk! Good choice! Michael *TITLE* Neural Networks and Bayes Rule *KEYWORDS* Deep Graphical Models, Reasoning with NNs *ABSTRACT* Relational or structured reasoning is an important current research challenge. The classical approach to this challenge is a templated graphical model: highly expressive, with well-founded semantics, but (at least naively) difficult to scale up, and difficult to combine with the most effective supervised learning methods. More recently, researchers have designed many different deep network architectures for structured reasoning problems, with almost the flip set of advantages and disadvantages. Can we get the best of both worlds? That is, can we design deep nets that look more like graphical models, or graphical models that look more like deep nets, so that we get a framework that is both practical and "semantic"? This talk will take a look at some progress toward such a hybrid framework. *BIO* Dr. Gordon is the Research Director of Microsoft Research Montreal. He is on leave as a Professor in the Department of Machine Learning at Carnegie Mellon University, where he has also served as Interim Department Head and as Associate Department Head for Education. His research interests include artificial intelligence, statistical machine learning, game theory, multi-robot systems, and planning in probabilistic, adversarial, and general-sum domains. His previous appointments include Visiting Professor at the Stanford Computer Science Department and Principal Scientist at Burning Glass Technologies in San Diego. Dr. Gordon received his B.A. in Computer Science from Cornell University in 1991, and his Ph.D. in Computer Science from Carnegie Mellon University in 1999.

2 3

No [Tea Talk] this week
by Michael Noukhovitch 27 Aoû '18

27 Aoû '18

This week's tea talk will be cancelled in lieu of the NLP Workshop <http://www.crm.umontreal.ca/2018/Langue18/progNLP_en.html>. To all those attending, have fun! Michael

1 0

[Tea Talk] David Rolnick (UPenn) August 24 2018 10:30 AA6214
by Michael Noukhovitch 24 Aoû '18

24 Aoû '18

Sorry for the delay but tomorrow we have *David Rolnick * visiting with Konrad Kording's lab at * UPenn * giving a talk on *August 24 2018* at *10:30* in room *AA6214* **Note the different room AA6214** Will this talk be streamed <https://mila.bluejeans.com/809027115/webrtc>? yes David will be free until only 1:30~2pm so if you'd like to go with him for lunch, stay after the talk and join in! Be sure to come and hear about the impactful, *profound* insights David has to share. Michael P.S. since David was so kind as to step in on short notice, I'll be making (or buying) cookies for early attendees ;) *TITLE* The impact of depth on expressivity and learning *KEYWORDS *Deep Learning Theory *ABSTRACT* Deeper networks are more powerful than shallow ones, but can be harder to train. In this talk, we will rigorously investigate why both of these statements are true. Specifically, we will prove that depth leads to an exponentially greater ability to express even simple polynomial functions. We will identify why some initializations and architectures impede learning in deeper networks, and demonstrate (both mathematically and empirically) several principles to bear in mind when designing an MLP/ConvNet/ResNet that will learn effectively. *BIO* David Rolnick completed his Ph.D. in Mathematics at MIT this year, co-advised by Nir Shavit, Max Tegmark, and Ed Boyden. His work focuses on the mathematical foundations of artificial and biological neural networks. A former NSF Graduate Research Fellow and Fulbright Scholar, David has also worked on machine learning research as an intern at Google and DeepMind. He will be joining Konrad Kording's group at UPenn as an NSF Mathematical Sciences Postdoctoral Research Fellow.

2 3

[Visitor Talk] Gael Varoquaux, Friday 2pm
by Pascal Vincent 23 Aoû '18

23 Aoû '18

Hello all, this Friday, we will have the visit of *Gaël Varoquaux* from INRIA, who will be giving a talk at 2 p.m. in AA-3195. Gaël is famous for creating the highly successful scikit-learn software and for his past work on brain imaging and neuroscience. But he will be talking primarily about his current research interests, which are machine learning approaches applicable to more general types of data. Xavier Bouthillier is coordinating his Friday visit, please get in touch with him (xavier.bouthillier(a)gmail.com) if you want to set aside time to meet the speaker. *When:* this Friday, august 24th, at 2 p.m. *Where:* Room 3195, Pavillon André Aisenstadt *Title:* Simple representations for learning: factorizations and similarities *Abstract* Real-life data seldom comes in the ideal form for statistical learning. This talk will focus on high-dimensional problems for signals and discrete entities: when dealing with many, correlated, signals or entities, it is useful to extract representations that capture these correlations. Matrix factorization models provide simple but powerful representations. They are used for recommender systems across discrete entities such as users and products, or to learn good dictionaries to represent images. However they entail large computing costs on very high-dimensional data, databases with many products or high-resolution images. I will present an algorithm to factorize huge matrices based on stochastic subsampling that gives up to 10-fold speed-ups [1]. With discrete entities, the explosion of dimensionality may be due to variations in how a smaller number of categories are represented. Such a problem of "dirty categories" is typical of uncurated data sources. I will discuss how encoding this data based on similarities recovers a useful category structure with no preprocessing. I will show how it interpolates between one-hot encoding and techniques used in character-level natural language processing. [1] Stochastic subsampling for factorizing huge matrices A Mensch, J Mairal, B Thirion, G Varoquaux IEEE Transactions on Signal Processing 66 (1), 113-128 [2] Similarity encoding for learning with dirty categorical variables. P Cerda, G Varoquaux, B Kégl Machine Learning (2018): 1-18

2 1

[Tea Talk] Abhishek Das (Georgia Tech) August 17 2018 10:30 AA3195
by Michael Noukhovitch 17 Aoû '18

17 Aoû '18

This week we have *Abhishek Das * from * Georgia Tech (and currently interning at FAIR Montreal) * giving a talk on *August 17 2018* at *10:30* in room *AA3195* Will this talk be streamed <https://mila.bluejeans.com/809027115/webrtc>? yes Want to meet the speaker? Sign up here <https://calendar.google.com/calendar/selfsched?sstoken=UVBVTVF0X25Nd09LfGRl…> ☕ talk ->🚶to 3195 Michael *TITLE* Connecting Vision and Language to Actions *KEYWORDS *language understanding, VQA, embodied agents *ABSTRACT* Building intelligent agents that possess the ability to perceive the rich visual environment around us, communicate this understanding in natural language to humans and other agents, and execute actions in a physical environment, is a long-term goal of Artificial Intelligence. In this talk, I will present some of my recent work at various points on this spectrum in connecting vision and language to actions; from Visual Dialog (CVPR17, ICCV17, HCOMP17) -- where we develop models capable of holding free-form visually-grounded natural language conversation towards a downstream goal and ways to evaluate them -- to Embodied Question Answering (CVPR18) -- where we augment these models to actively navigate in simulated environments and gather visual information necessary for answering questions. *BIO* Abhishek Das is a Computer Science PhD student at Georgia Institute of Technology, advised by Dhruv Batra, and working closely with Devi Parikh. He is interested in deep learning and its applications in building agents that can see (computer vision), talk (language modeling), act (reinforcement learning), and reason. He is a recipient of an Adobe Research Fellowship and a Snap Research Fellowship. He has held internship positions at Queensland Brain Institute (Winter 2013, Winter 2014), Virginia Tech (Fall 2015) and Facebook AI Research (Summer 2017, Winter 2018, Summer 2018). He graduated from Indian Institute of Technology Roorkee in 2015 with a Bachelor's degree in Electrical Engineering.

1 2

[Tea Talk] Wengong Jin (MIT) August 10 2018 10:30 AA3195
by Michael Noukhovitch 10 Aoû '18

10 Aoû '18

This week we have *Wengong Jin * from * MIT * giving a talk on *August 10 2018* at *10:30* in room *AA3195* Will this talk be streamed <https://mila.bluejeans.com/809027115/webrtc>? yes Wengong is free to meet until 1:30pm so email me if you'd be interested in grabbing lunch with him or talking with him before the talk. If you're interested in the field, I'm certain the talk will get a great reaction! Michael *TITLE* Graph Representation Learning for Chemistry *KEYWORDS *computational chemistry, graph learning, generative models *ABSTRACT* Learning representation for molecules and reactions is crucial for computational chemistry. Traditional methods, based on expert designed molecular fingerprints or reaction templates, have lead to suboptimal predictive performance and poor scalability. In this talk, I will present our recent effort in dealing with these challenges. Particularly, I will demonstrate two novel approaches, showing how graph neural network could overcome these challenges, and how domain knowledge could help us design better graph neural network architectures. In the first part of the talk, I will describe a graph convolutional network (with self-attention) for predicting organic reaction outcomes. Our model significantly outperforms traditional template-based methods, approaching expert chemist performance while running 100x faster. In the second part of the talk, I will present a generative model called Junction Tree Variational Autoencoder for molecular graph generation, and show how it could be used for drug discovery. *BIO* Wengong Jin is currently a PhD student in Computer Science and Artificial Intelligence Lab (CSAIL) at MIT, working with Prof. Regina Barzilay and Tommi Jaakkola. Before he finished his Master of Science degree from MIT in 2018, and Bachelor of Engineering from Shanghai Jiao Tong University in 2016. His research interest lies in machine learning, with particular focus on deep learning methods for graphs. His recent research on molecular graph generation has been covered as spotlight news in MIT homepage. He is also honored with 2018 Ho-Ching and Han-Ching scholarship.

1 1

Resting-state and Brain Connectivity 2018, in Montreal
by Amir Shmuel, Dr. 07 Aoû '18

07 Aoû '18

For Montreal’s machine learning experts who are interested in inspiration for their neural networks and/or applications to neuroscience: Resting State and Brain Connectivity 2018 in Montreal Educational course: September 24-25, 2018 Main conference: September 26-28, 2018 We have received over 285 submissions! Abstract submission will remain open for late submissions until 11:59 PM on August 11, 2018. Early registration fee applies until 11:59 PM on August 15, 2018. Amir Shmuel Chair of the organizing committee Resting-State and Brain Connectivity 2018 Montreal, QC, Canada http://restingstate.com/2018/ [http://restingstate.com/2018/wp-content/uploads/2017/06/resting-state-logo-…]

1 0

[Extra Tea Talk] Berton Earnshaw (Recursion Pharmaceuticals) Mon Aug 6 12PM AA3195
by Michael Noukhovitch 06 Aoû '18

06 Aoû '18

Early next week we have an extra talk by *Berton Earnshaw*, Senior Director of Data Science Research at *Recursion Pharmaceuticals* giving a talk on *Monday August 6 2018* at *12pm - 1pm* in room *AA3195.* This will be a lunch talk and mingle with *food served!* Will this talk be streamed <https://mila.bluejeans.com/809027115/webrtc>? Yes Want to meet with Berton in the afternoon? https://calendar.google.com/calendar/selfsched?sstoken=UVBVTVF0X25Nd09LfGRl… Come for the food and machine learning, stay because the talk will be awesome. Michael *TITLE *Representation learning in human cellular biology *KEYWORDS* bioinformatics, representation learning *ABSTRACT* The automated labs at Recursion Pharmaceuticals currently run more than 100,000 experiments every week, and each experiment involves imaging the morphology of human cells under various biological and chemical perturbations. The resulting data set is the largest and most diverse of its kind, and is used to train networks that learn to predict which compounds treat which diseases, and why. In this talk, I will describe the challenges of learning on biological image sets, discuss some approaches we have taken to learning representations of these images, and describe how we use these representations to discover candidate treatments for disease. *BIO * Berton earned a PhD in math from the University of Utah in 2007, where he designed biophysical models of protein trafficking at synapses during episodes of learning and memory formation. In 2010, after two post-docs at Utah and Michigan State University spent developing the theory of non-autonomous master equations, Berton took some time off of academia and co-founded the call center software company Perfect Pitch where he led as CTO. In 2012 he joined the founding team of Red Brain Labs as Director of Operations and Data Science, where he worked to optimize call centers using machine learning and simulation. In 2014, Red Brain Labs was acquired by Savvysherpa where, as a Senior Scientist and Principal, Berton led various research projects applying machine learning to insurance claims and medical records with the goal of improving healthcare outcomes. In 2017, Berton joined Recursion Pharmaceuticals as Senior Director of Data Science Research, where he leads the research team in applying deep learning to millions of images of human cells in search of treatments for hundreds of diseases. *COMPANY BIO* At Recursion Pharmaceuticals, we have raised over $80M to apply machine learning to one of the most unique datasets in existence - tens of millions of images of cells under a huge number of biological and chemical perturbations, generated in our own labs - in order to find treatments for hundreds of diseases. Our long term mission is to decode biology to radically improve lives - we want to understand biology so well that we can fix most things that go wrong in our bodies.

2 3

[Tea Talk] Petar Veličković (Cambridge / Mila) August 32018 10:30 AA3195
by Michael Noukhovitch 03 Aoû '18

03 Aoû '18

This week we have *Petar Veličković * from * Cambridge/Mila * giving a talk on *August 3rd 2018* at *10:30* in room *AA3195* Will this talk be streamed <https://mila.bluejeans.com/1365483656/webrtc>? yes Come to the talk and make sure to pay attention, because this talk should be great to connect the dots! Michael *TITLE* Keeping our graphs attentive *KEYWORDS *graph structured, attention, computational biology *ABSTRACT* A multitude of important real-world datasets (especially in biology) come together with some form of graph structure: social networks, citation networks, protein-protein interactions, brain connectome data, etc. Extending neural networks to be able to properly deal with this kind of data is therefore a very important direction for machine learning research, but one that has received comparatively rather low levels of attention until very recently. Attentional mechanisms represent a very promising direction for extending the established convolutional operator on images to work on arbitrary graphs, as they satisfy many of the desirable features for a convolutional operator. Through this talk, I will focus on my work on Graph Attention Networks (GATs), where these theoretical properties have been further validated by solid results on transductive as well as inductive node classification benchmarks. I will also outline some of the earlier efforts towards deploying attention-style operators on graph structures, as well as very exciting recent work that expands on GATs and deploys them in more general circumstances (such as EAGCN, DeepInf, and applications to solving the Travelling Salesman Problem). Time permitting, I will also present some of the relevant related graph-based work in the computational biology and medical imaging domains that I have been involved in. *BIO* Petar Veličković is currently a final-year PhD student in Machine Learning and Bioinformatics at the Department of Computer Science and Technology of the University of Cambridge. He also holds a BA degree in Computer Science from Cambridge, having completed the Computer Science Tripos in 2015. In addition, he has been involved in research placements at Nokia Bell Labs (working with Nicholas Lane) and the Montréal Institute of Learning Algorithms (working with Adriana Romero and Yoshua Bengio). His current research interests broadly involve devising neural network architectures that operate on nontrivially structured data (such as graphs), and their applications in bioinformatics and medicine. He has published his work in these areas at both machine learning venues (ICLR, NIPS ML4H, ICML WCB) and biomedical venues and journals (Bioinformatics, PervasiveHealth)

2 3

Aller à la page :

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

Lisa_seminaires