- Lisa_teatalk - mailman.iro.umontreal.ca

[Tea Talk] Bart Van Parys (MIT) Fri Jan 26 1:30PM AA1360
by Michael Noukhovitch 27 Jan '18

27 Jan '18

Hi all! We will have a researcher from MIT, *Bart Van Parys*, giving a talk on *Friday Jan 26th at 1:30PM* in room *AA1360*. Note the later time, *1:30!* This is a seminar is organized in collaboration with the Canada Research Chair in Decision Making under Uncertainty at GERAD! So this is ideal for all the optimization fans out there! Based on my experience, this will be a great talk and coming to it should be an easy decision! Michael *KEYWORDS *Optimization, Sparse Regression, Generalization *TITLE *Modern Optimization for Sparse Learning and Robust Analytics *ABSTRACT* We discuss the tremendous potential of integer optimization methods for learning predictive models from high-dimensional data via exact sparse regression. We show that novel integer formulations can solve exact sparse regression problems of sizes counting p=100,000s covariates for n=10,000s of samples. That is, two orders of magnitude more than current state of the art methods. We also indicate that robust optimization methods can help practitioners make data-driven decisions which are safeguarded against over-calibration to one particular data set. We claim that robust optimization methods have an enormous untapped potential when making subsequent decisions based on data. *BIO* Bart Van Parys is currently a postdoctoral researcher working with Prof. Dimitris Bertsimas at the MIT Sloan School of Management. His research interests are situated on the interface between optimization and machine learning. In 2015 he obtained his Ph.D. in control theory at the Swiss Federal Institute of Technology (ETH) in Zurich under the supervision of Prof. Manfred Morari. He received his M.E. from the University of Leuven in 2011.

1 1

[Tea Talk] Timothy O'Donnell (McGill) Fri Jan 19 10:30AM AA1360
by Michael Noukhovitch 19 Jan '18

19 Jan '18

*Note: *For the new year, we're moving to a different room, *AA1360* on the first floor! Exciting! For the first tea talk of the year we are going to have *Timothy O'Donnell*, professor at *McGill,* giving a talk on *Friday Jan 19th* at* 10:30AM* in room *AA1360*. This should be a real gronk of talk, and I'd encourage everyone to attend to it! Michael *KEYWORDS *Computational Linguistics, Language Learning, Linguistic Structure *TITLE* Computation, Storage, and Generalization in Language *ABSTRACT* A celebrated property of natural language is creativity, the ability to combine stored units to derive new expressions. This feature is found across multiple levels of linguistic structure: Sounds can be combined to form novel roots, suffixes, and affixes, which can be combined to form novel words, which can be combined to form novel sentences. However, each human language is characterized by its own inventory of units at each level and its own constraints on their combination. A syllable like "derp" sounds like a better morpheme of English than a syllable like "denp." The existence of the words "warmth" and "truth" do not imply the possibility of "coolth," but "warmness," "trueness," and "coolness" are all grammatical. And speakers of English will naturally drop the phrase "on the counter" from "John made dinner on the counter", but not from "John put dinner on the counter". How do learners acquire the inventory of units and constraints at each level of linguistic structure that are particular to their own language? I will discuss a theoretical framework designed to address this question and a number of specific applications. The approach is based on the idea that this problem can be approached by formulating a variety of linguistic learning problems as problems of program induction and then optimizing a tradeoff between a pressure to store fewer, more reusable primitive units and a pressure to account for each linguistic expression in as few computational steps as possible. Although the idea behind this tradeoff is an old one, it has surprisingly deep and far-reaching consequences when applied to domain-specific models of linguistic computation. I will show how this approach can shed light on a number of problems across the linguistic system. *BIO* Tim O'Donnell is an assistant professor in the department of linguistics at McGill University. Previously he was a research scientist at MIT in the department of Brain and Cognitive Sciences. His research focuses on developing mathematical and computational models of language learning and processing. His work draws on techniques from computational linguistics, and artificial intelligence integrating ideas from theoretical linguistics and methods from experimental psychology.

1 1

No [Tea Talk] this week
by Michael Noukhovitch 10 Jan '18

10 Jan '18

Happy new year and welcome back! This week's tea talk has been delayed til March so *no tea talk this week*. Instead you can come talk to me about tea for an hour, no scheduling needed! Cheers! Michael

1 0

Thesis proposal by Tristan Sylvain, Monday 9:30 @ 3195
by Yoshua Bengio 18 Déc '17

18 Déc '17

Hello, Tristan Sylvain will present his thesis proposal Monday 9:30-11:30 at PAA 3195. Everyone is welcome. Jury: President: Simon Lacoste-Julien Supervisor: Yoshua Bengio Member: Ioannis Mitliagkas Deep Neural Networks applied to medical data and Zero-Shot learning This presentation details two ongoing projects, applying deep learning to the problem of survival analysis for medical data, and using an adversarial approach to zero-shot learning. These very different problems have in common the fact that deep learning allows the learning of efficient representations of data that permit generalization when dealing with both highly noisy (in the medical case) and weakly supervised data (the zero-shot learning case). For the survival data, we used a multi-objective training procedure that allowed us to attain state of the art on many different public datasets, when compared to classical methods such as the Cox model. This presentation will provide both insights on the architecture used, and the medical context. Overall this problem is an interesting instance of applying deep learning to medical problems. Zero-shot learning is an active field of machine learning characterized by the fact that the classes we perform test time classification on are all unseen during the training procedure. To compensate for the lack of information on these classes, we are usually given a set of semantic attributes to allow some form of transfer from a model learned on seen train classes. This part of the presentation will present a current project aiming to use conditional generative adversarial networks to generate fake prototypes for unseen classes in a zero-shot learning context. I will present current results, and some of the challenges that lie ahead.

1 1

[Tea Talk] Romain Laroche (MS Maluuba) Fri Dec 1 10:30AM, AA6214
by Michael Noukhovitch 01 Déc '17

01 Déc '17

This week we have a researcher from MS Maluuba, *Romain Laroche,* giving a talk on *Friday Dec 1* at* 10:30AM* in room *AA6214*. See you there, this RL talk is sure to be rewarding! Michael *KEYWORDS* policy-based RL, bootstrapping, data/computational efficiency *TITLE *Safe Policy Improvement with Baseline Bootstrapping *ABSTRACT* A common goal in Reinforcement Learning is to derive a good strategy given a limited batch of data. In this paper, we propose a new strategy to compute a safe policy, guaranteed to perform at least as well as a given baseline strategy. We advocate that the assumptions made in previous work are too strong for real world applications and propose new algorithms allowing those assumptions to be satisfied only in a subset of the state-action pairs. While significantly relaxing the assumptions, our algorithms achieve the same accuracy guarantees than the previous work, and are also much more computationally efficient. We also show that the algorithms can be adapted to model-free Reinforcement Learning. *BIO* Graduated from Ecole Polytechnique in 2001 and then from Telecom PariTech in 2003, Romain Laroche joined the dialogue team at Orange in Paris, where he defended in 2010 a corporate PhD on Reinforcement Learning for industrial dialogue systems at Université Pierre et Marie Curie (Paris VI). Romain joined Maluuba in 2016, and is now a researcher at Microsoft Research Maluuba. During the past 6 years he supervised 5 PhDs, 3 postdocs, and a dozen of undergrad interns. Counting more than 40 papers at international conferences, his interest focuses now on Reinforcement Learning. His preferred application domains are dialogue systems (still his primary real-world motivation), Atari games (for benchmarking against other algorithms), and navigation toy problems (for empirical analysis and algorithm design).

1 1

[Tea Talk] Sergey Plis (MRN) Fri Nov 24, 10:30AM, AA6214
by Michael Noukhovitch 24 Nov '17

24 Nov '17

This week we have TWO tea talks (what a treat!) Apart from Martin Arjovsky's talk on Thursday, we also have the director of machine learning at the Mind Research Network, *Sergey Plis,* giving a talk on *Friday Nov 24* at* 10:30AM* in room *AA6214*. Hope this gives you cause to come, there's lots to learn! Michael *KEYWORDS *neuroscience, brain imaging, bayesian networks, structure learning *TITLE *Causal learning in complex time series datasets *ABSTRACT* Cortical neurons form coherent functional networks that are surprisingly stable across subjects and conditions. Together these networks comprise a set of functional units of the brain. Understanding their interactions can lead to better understanding of brain’s function and dysfunction due to disruption of the interaction structure. The most common way of assessing this structure are cross-correlation matrices (usually referred to as functional connectivity) but their drawbacks make us turn to modelling interactions via the directed graph of a Bayesian network (effective connectivity). Various brain imaging modalities contain different and arguably complementary information about interactions of functional network. Our goal is to bring together multimodal information to improve effective connectivity estimates. In particular, I focus on the problem of finding a common denominator for causal structures learned from time series at different time scales. I will demonstrate 1) a general theory which explains the effects of undersampling on apparent causal structure in terms of the true structure at the causal time scale; 2) a forward algorithm that computes a graph structure at any given undersampling rate; and 3) some inverse algorithms to compute all of the candidate graphs that could have generated the given undersampled structure. Although the problem is inspired by the questions from brain imaging, all time series models that are assessing structural relations are affected by the phenomena that I will describe. This includes Dynamic Bayesian Networks on the causal learning side and such models as vector autoregressive (VAR), structured VAR, Granger causality. *BASED ON* Cook, John W., David Danks, and Sergey M. Plis. (2017) "Learning Dynamic Structure from Undersampled Data <https://www.cs.purdue.edu/homes/eb/causal-uai17/papers/7.pdf>." UAI Causality Workshop A constraint optimization approach to causal discovery from subsampled time series data <https://scholar.google.ca/scholar?oi=bibs&cluster=5464303113063724661&btnI=…> A Hyttinen, S Plis, M Järvisalo, F Eberhardt, D Danks - International Journal of Approximate Reasoning, 2017 Danks, D., & Plis, S. M. (2017). Amalgamating evidence of dynamics. Synthese. [PDF <http://www.andrew.cmu.edu/user/ddanks/papers/Amalgamation-Synthese-Prepub.p…> ] Plis, S., Danks, D., Freeman, C., & Calhoun, V. (2015). Rate-agnostic (causal) structure learning. In Advances in neural information processing systems 28 (pp. 3303-3311). La Jolla, CA: The NIPS Foundation. (First two authors contributed equally.) [PDF <http://www.andrew.cmu.edu/user/ddanks/papers/RASL-NIPS-final.pdf>] Plis, S., Danks, D., & Yang, J. (2015). Mesochronal structure learning. In M. Meila & T. Heskes (Eds.), Uncertainty in artificial intelligence 31 (UAI-2015) (pp. 702-711). Corvallis, OR: AUAI Press. (First two authors contributed equally.) [PDF <http://www.andrew.cmu.edu/user/ddanks/papers/MSL-UAI15-final.pdf>] Danks, D., & Plis, S. (2014). Learning causal structure from undersampled time series. In JMLR: Workshop and Conference Proceedings. (First two authors contributed equally.) [PDF <http://www.andrew.cmu.edu/user/ddanks/papers/DanksPlis-Final.pdf>] *BIO*Dr. Sergey Plis is a Director of Machine Learning at the Mind Research Network. His research interests lie in developing novel and applying existing techniques and approaches to analyzing large scale datasets in multimodal brain imaging and other domains. He develops tools that fall within the fields of machine learning and data science. One key goal is to take advantage of the strengths of imaging modalities and infer structure and patterns that are hard to obtain non-invasively and/or that are unavailable for direct observation. In the long term this amounts to developing methods capable of revealing mechanisms used by the brain for forming task specific transient interaction networks and their cognition-inducing interactions via multimodal fusion at features and interaction levels. Ongoing work is focused on inferring multimodal probabilistic and causal descriptions of these function-induced networks based on fusion of fast and slow imaging modalities. This includes feature estimation via deep learning-based pattern recognition and learning causal graphical models. Dr. Plis is also the R&D director at a young machine learning startup reason8.ai

1 1

[Tea Talk] Martin Arjovsky (NYU) Thr Nov 23, 10:30AM, AA6214
by Michael Noukhovitch 23 Nov '17

23 Nov '17

This week we have a former MILA intern, *Martin Arjovsky,* now at NYU giving a talk on *Thursday Nov 23* at* 10:30AM* in room *AA6214*. Be there or be probabilistically sqaure! Michael *KEYWORDS* deep learning theory, probability theory, geometry *TITLE* A few geometrical insights into unsupervised learning *ABSTRACT* This short talk will address the connections and properties of the energy (including Cramer), and Wasserstein distances, and how they connect with Maximum Mean Discrepancy. We will dive into the similarities and differences among them, and how this impacts their behaviour in different ways. Some of the results are a contextualization of old studies, and some will be presented in this talk for the first time. A recurrent theme in the talk will be the study of the geometry of the space of probability distributions, and how this might shed clarity on the behaviour of different algorithms in practice. *BIO* I'm Martin Arjovsky, I'm currently doing my PhD at New York University, and being advised by Léon Bottou. I did my undergraduate and master's in the University of Buenos Aires, Argentina (my home country). In the middle I took a year off to do internships in different places (Google, Facebook, Microsoft, and the Université de Montréal). My master's thesis advisor was Yoshua Bengio, who also advised me during my stay at UdeM. In general I'm interested in the intersection between learning and mathematics, how we can ground the different learning processes that are involved in different problems, and leverage this knowledge to develop better algorithms.

1 1

[Tea Talk] Zac Kenton (MILA) Fri Nov 17, 10:30AM, AA6214
by Michael Noukhovitch 17 Nov '17

17 Nov '17

This week we have our very own *Zac Kenton*, a visiting researcher at MILA giving a talk on *Friday Nov 17* at* 10:30AM* in room *AA6214*. I've suggested that the full title should be "Three factors influencing minima in SGD: You'll never believe #1 and #3!" See you there! Michael *KEYWORDS:* SGD, Deep Learning Theory, Generalization *TITLE* Three factors influencing minima in SGD *ABSTRACT* We focus on the importance of noise in stochastic gradient descent (SGD) based training of deep neural networks (DNNs). We develop theory that studies SGD training as a stochastic differential equation and show that its stationary distribution is related to the loss surface. Our analysis suggests that the combination of batch size, learning rate, and the variance of the true loss gradients acts as a hyper- parameter steering the behavior of SGD and determines the trade-offs between the depth and width of the minima that SGD converges to. In a nutshell, a higher ratio of learning rate to batch size leads to wider minima. We validate our theory by examining the correlation between these three factors and the final performance and sharpness of the minimum found. As a verification of our theory, we empirically demonstrate that the learning dynamics is similar between experiments with different learning rates and batch sizes in SGD if the ratio of learning rate to batch size is the same. *BIO* Zac studied Mathematics at the University of Cambridge for a bachelors and masters 2009-2013. He then completed a PhD in theoretical physics at the Centre for Research in String Theory, Queen Mary University of London, 2013- August 2017. His thesis was on string theory and early universe inflationary cosmology. In the final stages of his PhD he also worked as a data scientist at ASI Data Science, a London-based data science startup. At MILA he's been working with Stanislaw Jastrzebski, Devansh Arpit and Prof Bengio on topics around generalization in SGD.

1 1

Re: [Lisa_teatalk] [Lisa_labo] please sign letter to ask government to work against lethal autonomous weapons at the UN
by Yoshua Bengio 15 Nov '17

15 Nov '17

Please do! On Nov 10, 2017 09:28, "Joseph Paul Cohen" <joseph(a)josephpcohen.com> wrote: Can I forward this email to researchers and ask them to sign the petition? -Joseph On Nov 10, 2017 06:57, "Yoshua Bengio" <yoshua.umontreal(a)gmail.com> wrote: Hello, Especially if you are an AI researcher in Canada, please consider and sign this letter asking the Canadian government to urgently address the challenge of lethal autonomous weapons (often called “killer robots”) and to take a leading position against Autonomous Weapon Systems on the international stage at the upcoming UN meetings in Geneva. The leading signatories of this letter are Ian Kerr <https://www.facebook.com/ian.kerr?fref=mentions>, Geoff Hinton, Richard S. Sutton <https://www.facebook.com/richard.s.sutton.7?fref=mentions>, Doina Precup <https://www.facebook.com/doina.precup?fref=mentions> and myself. Let's work together towards a socially responsible and ethically grounded use of AI. https://techlaw.uottawa.ca/bankillerai -- Yoshua _______________________________________________ Lisa_labo mailing list Lisa_labo(a)iro.umontreal.ca https://webmail.iro.umontreal.ca/mailman/listinfo/lisa_labo

2 1

[Tea Talk] Rosemary Ke (MILA) Nov 10 10:30AM AA6214
by Michael Noukhovitch 10 Nov '17

10 Nov '17

This week we have our very own *Rosemary Ke* giving a talk on *Friday Nov 10* at* 10:30AM* in room *AA6214*. See you there! Michael *KEYWORDS* RNN, BPTT, Attention Mechanism *TITLE *Sparse Attentive Backtracking: Long-Range Credit Assignment in Recurrent Networks *ABSTRACT* A major drawback of backpropagation through time (BPTT) is the difficulty of learning long-term dependencies, coming from having to propagate credit information backwards through every single step of the forward computation. This makes BPTT both computationally impractical and biologically implausible. For this reason, full backpropagation through time is rarely used on long sequences, and truncated backpropagation through time is used as a heuristic. However, this usually leads to biased estimates of the gradient in which longer term dependencies are ignored. Addressing this issue, we propose an alternative algorithm, Sparse Attentive Backtracking, which might also be related to principles used by brains to learn long-term dependencies. Sparse Attentive Backtracking learns an attention mechanism over the hidden states of the past and selectively backpropagates through paths with high attention weights. This allows the model to learn long term dependencies while only backtracking for a small number of time steps, not just from the recent past but also from attended relevant past states.

1 1