Date( Winter 2014 ) Category Seminar Info 2014/04/15 Graduate Seminar Series Place: MC103 Time: 12:10 - 12:40 Speaker: Pierre-Luc Bacon Affiliation: PhD Student, SOCS Title: A short tutorial on starcluster and ipython parallel Abstract: I recently faced the challenge of computing some memory-hungry statistics on our modest lab machines. With a conference deadline approaching, I had to quickly find a solution that would not involve buying new hardware. Due to the administrative delays, applying to CLUMEQ was also not an option. I first experimented with a duck-tape homemade cluster solution that allowed me to pool the computing power of all our lab machines simultaneously. While fully functional, this solution was limited by the lack of administrative privileges on the SOCS machines. And then came Starcluster. I will tell you the story of how I used Amazon EC2 to fit 750Gb of data in RAM while sipping a Cappuccino from a fancy coffee place. Not only did I get my results back in less than 30 mins, but the overall experiment also only cost \$5. I will give a brief tutorial on how to set up Starcluster ( http://star.mit.edu/cluster/) with with IPython parallel. I will also show you how to reduce the cost of your experiment even more by bidding on cheap "spot instances". Roboticists, biologists, machine learnists, and all other "ists" dealing with abundant data are welcome. 2014/04/08 Graduate Seminar Series Place: MC103 Time: 12:10 - 12:40 Speaker: Bundit Laekhanukit Affiliation: PhD Student, McGill SOCS Title: The Impossibility of PAC-learning DFAs Abstract: Probably approximately correct learning (PAC learning) is a learning model proposed by Leslie Valiant that introduces the concepts of computational complexity to machine learning. In particular, given positive and negative samples, the learner is expected to find a polynomial-time computable function that could approximately distinguish between positive and negative samples. In this work, we study the PAC-learnability of DFAs, i.e., the case where positive and negative samples are drawn from an unknown DFA. We show that, unless NP=RP, DFAs are not PAC-learnable, thus, answering an open question raised 25 years ago by Pitt and Warmuth. This is a joint work with Parinya Chalermsook and Danupon Nanongkai. 2014/04/02 Graduate Seminar Series Place: MC320 Time: 12:10 - 12:40 Speaker: Annie Ying Affiliation: PhD Candidate, McGill SOCS Title: Tips on Applying for Ethics Approval for Research Involving Human Participants Abstract: For some of us who do research involving human participants, one of the first steps is to apply for approval from the McGill Reseasrch Ethics Board. In this talk, I will share some of my experience in writing the application for ethics approval in my research. 2014/03/25 Graduate Seminar Series Place: MC103 Time: 12:10 - 12:40 Speaker: Ouais Alsharif Affiliation: Title: Lifelong Learning of Discriminative Representations Abstract: We envision a machine learning service provider facing a continuous stream of problems with the same input domain, but with output domains that may differ. Clients present the provider with problems implicitly, by labeling a few example inputs, and then ask the provider to train models which reasonably extend their labelings to novel inputs. The provider wants to avoid constraining its users to a set of common labels, so it does not assume any particular correspondence between labels for a new task and labels for previously encountered tasks. To perform well in this setting, the provider needs a representation of the input domain which, in expectation, permits effective models for new problems to be learned efficiently from a small number of examples. While this bears a resemblance to settings considered in previous work on multitask and lifelong learning, our non-assumption of inter-task label correspondence leads to a novel algorithm: Lifelong Learner of Discriminative Representations (LLDR), which explicitly minimizes a proxy for the intra-task small-sample generalization error. We examine the relative benefits of our approach on a diverse set of real-world datasets in three significant scenarios: representation learning, multitask learning and lifelong learning. 2014/02/25 Faculty Candidate Talk Place: MC 103 Time: 9:30 - 10:30 Speaker: Finale Doshi-Velez Affiliation: Harvard School of Engineering and Applied Sciences and Harvard Medical School Area: Datamining and health informatics Title: Prediction and Interpretation with Latent Variable Models Abstract: Latent variable models provide a powerful tool for summarizing data through a set of hidden variables. These models are generally trained to maximize prediction accuracy, and modern latent variable models now do an excellent job of finding compact summaries of the data with high predictive power. However, there are many situations in which good predictions alone are not sufficient. Whether the hidden variables have inherent value by providing insights about the data, or whether we wish to interface with domain expert on how to improve a system, understanding what the discovered hidden variables mean is an important first step. In this talk, I will discuss how the language of probabilistic modeling naturally and flexibly allows us to incorporate information about how humans organize knowledge in addition to finding predictive summaries of data. In particular, I will talk about how a new model, GraphSparse LDA, discovers interpretable latent structures without sacrificing (and sometimes improving upon!) prediction accuracy. The model incorporates knowledge about the relationships between observed dimensions into a probabilistic framework to find a small set of human-interpretable "concepts" that summarize the observed data. This approach allows us to recover interpretable descriptions of novel, clinically-relevant autism subtypes from a medical data-set with thousands of dimensions. Biography of Speaker: Finale Doshi-Velez is a postdoctoral fellow jointly between Harvard's School of Engineering and Applied Sciences and the Center for Biomedical Informatics. She received her MSc from the University of Cambridge in 2009 (as a Marshall Scholar) and her PhD from MIT in 2012. She was selected as one of IEEE's "AI 10 to Watch" in 2013. Her research interests include latent variable modeling, sequential decision-making, and clinical applications. 2014/02/24 Graduate Seminar Series Place: MC320 Time: 12:10 - 12:40 Speaker: Annie Ying Affiliation: PhD Candidate, SOCS Title: Selection and Presentation Practices in Code Example Summarization Abstract: Code examples are an important source for answering questions about software libraries and applications. Many usage contexts for code examples require them to be distilled to their essence: for example when serving as cues to longer documents, or for reminding developers of a previously-known idiom. We conducted a study on the practices employed by 16 programmers to summarize code examples. As part of the study, we collected 156 pairs of code examples and their summaries, along with over 26 hours of think-aloud verbalizations detailing the decisions of the participants during their summarization activities. We report on the summarization process, provide a list of practices followed by the participants to summarize code examples, and propose empirically-supported hypotheses justifying the use of specific practices. The results provide a grounded basis for the development of code example summarization technology. Special acknowledgements to the study's participants, many of whom are from SOCS! 2014/02/20 Graduate Seminar Series Place: MC103 Time: 12:10 - 12:55 Speaker: Anqi Xu Affiliation: PhD student, McGill CIM Title: An Introduction to Open-Source 3D Printing Abstract: 3-D printing technologies have existed in research and industrial domains for several decades, although in recent years many of these designs and concepts are beginning to transition into the consumer and hobbyist communities. In this talk I will present a brief introduction to this exciting and fast-evolving field, and highlight a number of open-source (and other free) tools and services for 3-D printing hobbyists. I will also give a live demonstration of the end-to-end process of 3-D printing, starting from the design of a object model from scratch, to its preparation for 3-D printing, and culminating with an actual printed object by the end of the talk. Come check out various 3-D printed models and functional objects, learn about some free and powerful tools and services, and discover how YOU can participate in the open-source 3-D printing revolution! 2014/02/20 Faculty Candidate Talk Place: MC 437 Time: 9:30 - 10:30 Speaker: Julian McAuley Affiliation: Postdoc at Stanford University Area: Datamining and health informatics Title: Leveraging Data Across Time and Space to Build Predictive Models for Healthcare-Associated Infections Abstract: The proliferation of user-generated content on the web provides a wealth of opportunity to study humans through their online traces. I will discuss three aspects of my research, which aims to model and understand people's behavior online. First, I will develop rich models of opinions by combining structured data (such as ratings) with unstructured data (such as text). Second, I will describe how preferences and behavior evolve over time, in order to characterize the process by which people "acquire tastes" for products such as beer and wine. Finally, I will discuss how people organize their personal social networks into communities with common interests and interactions. These lines of research require models that are capable of handling high-dimensional, interdependent, and time-evolving data, in order to gain insights into how humans behave. Biography of Speaker: Julian McAuley is a postdoctoral scholar at Stanford University, where he works with Jure Leskovec on modeling the structure and dynamics of social networks. His current work is concerned with modeling opinions and behavior in online communities, especially with respect to their linguistic and temporal dimensions. Previously, Julian received his PhD from the ANU under Tiberio Caetano, with whom he worked on inference and learning in structured output spaces. His work has been featured in Time, Forbes, New Scientist, and Wired, and has received over 30,000 "likes" on Facebook. 2014/02/18 Faculty Candidate Talk Place: MC103 Time: 9:30 - 10:30 Speaker: Jenna Wiens Affiliation: MIT Area: Machine Learning and Data Mining Title: Leveraging Data Across Time and Space to Build Predictive Models for Healthcare-Associated Infections Abstract: The proliferation of electronic medical records holds out the promise of using machine learning and data mining to build models that will help healthcare providers improve patient outcomes. However, building useful models from these datasets presents many technical problems. The task is made challenging by the large number of factors, both intrinsic and extrinsic, influencing a patient’s risk of an adverse outcome, the inherent evolution of that risk over time, and the relative rarity of adverse outcomes. In this talk, I will describe the development and validation of hospital-specific models for predicting healthcare-associated infections (HAIs), one of the top-ten contributors to death in the US. I will show how by adapting techniques from time-series classification, transfer learning and multi-task learning one can learn a more accurate model for patient risk stratification for the HAI Clostridium difficile (C. diff). Applied to a held-out validation set of 25,000 patient admissions, our model achieved an area under the receiver operating characteristic curve of 0.81 (95%CI 0.78-0.84). On average, we can identify high-risk patients five days in advance of a positive test result. The model has been successfully integrated into the health record system at a large hospital in the US, and is being used to produce daily risk estimates for each in-patient. Clinicians at the hospital are now considering ways in which that information can be used to reduce the incidence of HAIs Biography of Speaker: Jenna Wiens is a Ph.D. Candidate in the Department of Electrical Engineering and Computer Science at the Massachusetts Institute of Technology (MIT). She holds an S.M. degree in EECS from MIT. She is interested in solving the technical challenges that arise when considering the practical application of machine learning in medicine. In addition to her work on predicting healthcare associated infections she has applied machine-learning methods to the automated interpretation of electrocardiograms and the extraction of strategically useful information from player tracking data in the NBA. 2014/02/14 Faculty Candidate Talk Place: MC103 Time: 10:30 - 11:30 Speaker: Jackie Chi Kit Cheung Affiliation: University of Toronto Area: Distributional Semantics Title: Towards Large-Scale Natural Language Inference with Distributional Semantics Abstract: Language understanding and semantic inference are crucial for solving complex natural language applications, from intelligent personal assistants to automatic summarization systems. However, current systems often require hand-coded information about the domain of interest, an approach that will not scale up to the large array of possible domains and topics in text collections today. In this talk, I demonstrate the potential of distributional semantics (DS), an approach to modeling meaning by using the contexts in which a word or phrase appears, to assist in acquiring domain knowledge and to support the desired inference. I present a method that integrates phrasal DS representations into a probabilistic model in order to learn about the important events and slots in a domain, resulting in state-of-the-art performance on template induction and multi-document summarization for systems that do not rely on hand-coded domain knowledge. I also propose to evaluate DS by their ability to support inference, the hallmark of any semantic formalism. These results demonstrate that the utility of DS for current natural language applications, and provide a principled framework for measuring progress towards automated inference in any domain. Biography of Speaker: Jackie CK Cheung is a PhD candidate at the University of Toronto. His research interests span several areas of natural language processing, including computational semantics, automatic summarization, and natural language generation. His work is supported by the Natural Sciences and Engineering Research Council of Canada (NSERC), as well as a Facebook Fellowship 2014/02/13 Graduate Seminar Series Place: MC103 Time: 12:10 - 12:40 Speaker: Hang Ma Affiliation: MSc Candidate, SOCS Title: Information Gathering and Reward Exploitation of Subgoals for POMDPs Abstract: Planning in large partially observable Markov decision processes (POMDPs) is challenging especially when a long planning horizon is required. A few recent algorithms successfully tackle this case but at the expense of weaker information-gathering capacity. In this paper, we propose \emph{Information Gathering and Reward Exploitation of Subgoals} (IGRES), a randomized POMDP planning algorithm that leverages information in the state space to automatically generate macro-actions'' that can tackle tasks with long planning horizons, while locally exploring the belief space to allow effective information gathering. Experimental results show that IGRES is an effective multi-purpose POMDP solver, providing state-of-the-art performance for both long horizon planning tasks and information gathering tasks on benchmark domains. Additional experiments with an ecological adaptive management problem presented in the IJCAI 2013 data challenge track indicate that IGRES is a promising tool for POMDP planning in real-world settings. 2014/02/11 Faculty Candidate Talk Place: MC 103 Time: 9:15 - 10:15 Speaker: Byron Wallace Affiliation: Brown University Area: Machine Learning in Evidence-Based Medicine Title: Machine Learning in Evidence-Based Medicine: Taming the Clinical Data Deluge Abstract: An unprecedented volume of biomedical evidence is being published today. Indeed, PubMed (a search engine for biomedical literature) now indexes more than 600,000 publications describing human clinical trials and upwards of 22 million articles in total. This volume of literature imposes a substantial burden on practitioners of Evidence-Based Medicine (EBM), which now informs all levels of healthcare. Systematic reviews are the cornerstone of EBM. They address a well-formulated clinical question by synthesizing the entirety of the published relevant evidence. To realize this aim, researchers must painstakingly identify the few tens of relevant articles among the hundreds of thousands of published clinical trials. Further exacerbating the situation, the cost of overlooking relevant articles is high: it is imperative that all relevant evidence is included in a synthesis, else the validity of the review is compromised. As reviews have become more complex and the literature base has exploded in volume, the evidence identification step has consumed an increasingly unsustainable amount of time. It is not uncommon for researchers to read tens of thousands of abstracts for a single review. If we are to realistically realize the promise of EBM (i.e., inform patient care with the best available evidence), we must develop computational methods to optimize the systematic review process. To this end, I will present novel data mining and machine learning methods that look to semi-automate the process of relevant literature discovery for EBM. These methods address the thorny properties inherent to the systematic review scenario (and indeed, to many tasks in health informatics). Specifically, these include: class imbalance and asymmetric costs; expensive and highly skilled domain experts with limited time resources; and multiple annotators of varying skill and price. In this talk I will address these issues in turn. In particular, I will present new perspectives on class imbalance, novel methods for exploiting dual supervision (i.e., labels on both instances and features), and new active learning techniques that address issues inherent to real-world applications (e.g., exploiting multiple experts in tandem). I will present results that demonstrate that these methods can reduce by half the workload involved in identifying relevant literature for systematic reviews, without sacrificing comprehensiveness. Finally, I will conclude by highlighting emerging and future work on automating next steps in the systematic review pipeline, and methods for making sense of biomedical data more generally. Biography of Speaker: Byron Wallace is an assistant research professor in the Department of Health Services, Policy & Practice at Brown University; he is also affiliated with the Brown Laboratory for Linguistic Processing (BLLIP) in the department of Computer Science. His research is in machine learning/data mining and natural language processing, with an emphasis on applications in health. Before moving to Brown, he completed his PhD in Computer Science at Tufts under the supervision of Carla Brodley. He was selected as the runner-up for the 2013 ACM SIGKDD Doctoral Dissertation Award and he was awarded the Tufts Outstanding Graduate Researcher at the Doctoral Level award in 2012 for his thesis work. 2014/02/06 Graduate Seminar Series Place: MC103 Time: 12:10 - 12:40 Speaker: Francisco Ferreira Affiliation: PhD student, McGill SOCS Area: Computation and Logic Group Title: An opinionated history of functional programming. Abstract: The opinionated nature of this talk is because of the selection of ideas. Functional programming has a very long history, and as such this presentation will revolve around only a handful of the ideas that motivate the current programming languages in this space. Topics range from the lambda calculus to Lisp, from logic to types, and how the paradigm helps with important topics like abstraction, and hiding implementation details. The talk will not cover the implementation techniques, or discuss the exhaustive family tree of languages; rather, the main point will be to briefly discuss some key insights that lead to languages such as Lisp, Scheme, Haskell, ML and Agda. 2014/01/23 Graduate Seminar Series Place: MC103 Time: 12:10 - 12:40 Speaker: Sheldon Andrews Affiliation: PhD Candidate, SOCS Title: FORKS: Interactive Compliant Mechanisms with Parallel State Computation Abstract: We present a method for the simulation of compliant, articulated structures using a plausible approximate model that focuses on modeling endpoint interaction. We approximate the structure’s behavior about a reference configuration, resulting in a first order reduced compliant system, or FORK-1S. Several levels of approximation are available depending on which parts and surfaces we would like to have interactive contact forces, allowing various levels of detail to be selected. Our approach is fast and computation of the full structure’s state may be parallelized. Our approach is suitable for stiff, articulate grippers, such as those used in robotic simulation, or physics based characters under static proportional derivative control. We demonstrate that simulations with our method can deal with kinematic chains and loops with non-uniform stiffness across joints, and that it produces plausible effects due to stiffness, damping, and inertia. http://www.cs.mcgill.ca/~sandre17/forks/ 2014/01/14 Graduate Seminar Series Place: MC103 Time: 12:10 - 12:40 Speaker: Christopher Dragert Affiliation: PhD Candidate, Games Research @ McGill (gr@m) lab Title: Exploring Model-Driven Development of Game AI Abstract: Game development is often an ad-hoc affair, focused more on performance and meeting deadlines than on writing maintainable and well-structured code. AI in particular is often highly customized to a specific game context, impairing reusability and verification. During my Ph.D. research, I explored how a model-driven approach to game AI can improve the development process. In this talk, I will present some of my major results, including the layered statechart-based AI formalism, which represents AI logic and behaviour in a fundamentally modular fashion. This approach leads naturally to effective reuse of AI behaviours, generative approaches to create varied NPC populations, tool support for managing development of game AI, and verification of AI behaviours through model-checking.