Coffee, Bars, Yogurt, Fruit, etc.
The Midwest ML Symposium aims to convene regional machine learning researchers for stimulating discussions and debates, to foster cross-institutional collaboration, and to showcase the collective talent of ML researchers at all career stages. [past events]
Where: Memorial Union@UW-Madsion
Contact: For any questions or concerns, please write to firstname.lastname@example.org
Parking: Visitor parking is available around UW-Madison Campus. Please check here for the availability and the directions: [parking info link]
Directions inside Memorial Union: The poster session will be in Tripp Commons on the second floor, while the other activities will be in the Great Hall on the fourth floor. Signs will be provided to guide participants to the rooms. Please check the [Map of the Memorial Union] for details.
The Midwest ML Symposium offers various opportunities of exposure. In addition to the satisfaction of supporting the regional Machine Learning community, you will be gratefully recognized in various media and materials and have the possibility to more closely engage with the participants.
Contact Information: Sponsors are encouraged to contact the Midwest ML Symposium organizing committee and to apply through the sponsorship application form. To discuss special requirements and to ask general questions regarding sponsorship of the Symposium, please contact us by email at: email@example.com
Rob Nowak (chair, UW-Madison), Maxim Raginsky (UIUC), Laura Balzano (UMich), Mikhail Belkin (OSU), Avrim Blum (TTI-C), Rebecca Willett (UChicago), Nati Srebro (TTI-C), Po-Ling Loh (UW-Madison), Matus Telgarsky (UIUC), Mike Franklin (UChicago).
University of Michigan, Ann Arbor
University of Wisconsin-Madison
University of Michigan, Ann Arbor
University of Illinois at Urbana-Champaign
The Ohio State University
University of Illinois at Urbana-Champaign
University of Wisconsin-Madison
University of Michigan, Ann Arbor
University of Minnesota
Check here for the list of posters. The poster session will be in Tripp Commons on the second floor, while the other activities will be in the Great Hall on the fourth floor.
Coffee, Bars, Yogurt, Fruit, etc.
ABSTRACT: Computer vision has seen major success in learning to recognize objects from massive “disembodied” Web photo collections labeled by human annotators. Yet cognitive science tells us that perception develops in the context of acting the world---and without intensive supervision. Meanwhile, many realistic vision tasks require not only categorizing a well-composed human-taken photo, but also actively deciding where to look in the first place. In the context of these challenges, we are exploring how machine perception benefits from anticipating the sights and sounds an agent will experience as a function of its own actions. Based on this premise, we introduce methods for learning to look around intelligently in novel environments, learning from video how to interact with objects, and perceiving audio-visual streams for both semantic and spatial context. Together, these are steps towards first-person perception, where interaction with the world is itself a supervisory signal.
BIO: Kristen Grauman is a Professor in the Department of Computer Science at the University of Texas at Austin and a Research Scientist at Facebook AI Research. Her research in computer vision and machine learning focuses on visual recognition and search. Before joining UT Austin in 2007, she received her Ph.D. at MIT. She is a AAAI Fellow, a Sloan Fellow, and a recipient of the NSF CAREER, ONR YIP, PECASE, PAMI Young Researcher award, and the 2013 IJCAI Computers and Thought Award. She and her collaborators were recognized with best paper awards at CVPR 2008, ICCV 2011, ACCV 2016, and a 2017 Helmholtz Prize “test of time” award. She served as a Program Chair of the Conference on Computer Vision and Pattern Recognition (CVPR) in 2015 and Neural Information Processing Systems (NeurIPS) in 2018, and she currently serves as Associate Editor-in-Chief for the Transactions on Pattern Analysis and Machine Intelligence (PAMI).
Understanding Learned Models by Identifying Important Features at the Right Resolution, Akshay Sood
Random sampling and efficient algorithms for multiscale PDEs, Ke Chen
Learning to Control Renewal Processes with Bandit Feedback, Semih Cayci
Speeding up Distributed Computing through Coding, Konstantinos Konstantinidis
On the Optimal Risk and Optimal Classifier in the Presence of an Adversary, Muni Pydi
All Nearest Neighbors from Noisy Distances, Blake Mason
Substituting ReLUs with Hermite Polynomials gives faster convergence for SSL, Sathya Ravi
Learning to Solve Inverse Problems with Neumann Networks, Davis Gilton
Target-Based Temporal-Difference Learning, Donghwan Lee
Convergence and Margin of Adversarial Training on Linearly Separable Data, Zachary Charles
ABSTRACT: Policy optimization (with neural networks as actor and critic) is the workhorse behind the success of deep reinforcement learning. However, its global convergence remains less understood, even in classical settings with linear function approximators. In this talk, I will show that coupled with neural networks, a variant of proximal/trust-region policy optimization (PPO/TRPO) globally converges to the optimal policy. In particular, I will illustrate how the overparametrization of neural networks enable us to establish strong guarantees. (Joint work with Qi Cai, Jason Lee, Boyi Liu, Zhuoran Yang)
BIO: Zhaoran Wang is an assistant professor at Northwestern University, working at the interface of machine learning, statistics, and optimization. He is the recipient of the AISTATS (Artificial Intelligence and Statistics Conference) notable paper award, ASA (American Statistical Association) best student paper in statistical learning and data mining, INFORMS (Institute for Operations Research and the Management Sciences) best student paper finalist in data mining, and the Microsoft fellowship.
ABSTRACT: Recent advances in deep reinforcement learning algorithms have shown great potential and success for solving many challenging problems, including Go game and robotic applications. Usually, these algorithms need a carefully designed reward function to guide training. However, in the real world, it is often non-trivial to design such a reward function, and sometimes the only signal available is usually obtained at the end of a trajectory and very expensive to obtain. To improve data efficiency, we hope to extract extra helpful information from past experience. In this talk, I will introduce several useful heuristics to learn from past experience to facilitate better policy training. In particular, I will briefly introduce 1) learning from success by self imitation, 2) learning with valuable hindsight goals, and 3) learning with rewards via credit assignment with sequence modeling.
BIO: Jian Peng has been an assistant professor of computer science at UIUC since 2015. His research interests include bioinformatics, cheminformatics and machine learning. Recently, Jian has received an NSF CAREER Award, a Pharma Foundation Award, and an Alfred P. Sloan Research Fellowship.
ABSTRACT: 3M’s legacy for developing unique products and manufacturing them efficiently around the world gives a massive platform to apply AI solutions to “real products” for an impressive practical success. In this talk, I will present several projects to emphasize how 3M is leveraging Artificial intelligence for improving the world today.
ABSTRACT: Machine learning has gained increasing attention and started to impact healthcare. In this talk, I will discuss the challenges and our work on using machine learning methods in mental health. I will discuss recent innovations on data, experiments, and models. As an example, I will elaborate findings that decipher the neurobehavioral signature of autism. I will then demonstrate deep learning models that are able to learn semantic attributes from complex natural scenes, leading to breakthrough performance in behavioral prediction and identifying people with autism. I will discuss these technological innovations in the context of clinical applications.
BIO: Catherine Qi Zhao is an assistant professor in the Department of Computer Science and Engineering at the University of Minnesota, Twin Cities. Her main research interests include computer vision, machine learning, cognitive neuroscience, and mental disorders. Dr. Zhao has published about 50 journal and conference papers in top-tier venues including Neuron, Current Biology, Nature Communications, TPAMI, IJCV, CVPR, ICCV, ECCV, NIPS and ICML, and edited a book with Springer, titled Computational and Cognitive Neuroscience of Vision, that provides a systematic and comprehensive overview of vision from various perspectives, ranging from neuroscience to cognition, and from computational principles to engineering developments.
ABSTRACT: Though the potential impact of machine learning in healthcare warrants genuine enthusiasm, the increasing computerization of the field is still often seen as a negative rather than a positive. The limited adoption of machine learning in healthcare to date points to the fact that there remain important challenges. In this talk, I will highlight two key challenges related to applying machine learning in healthcare: i) interpretability and ii) small sample size. First, machine learning has often been criticized for producing ‘black boxes.’ In this talk, I will argue that interpretability is neither necessary nor sufficient, demonstrating that even interpretable models can lack common sense. To address this issue, we propose a novel regularization method that enables the incorporation of domain knowledge during model training, leading to increased robustness. Second, machine learning techniques benefit from large amounts of data. However, oftentimes in healthcare we find ourselves in data poor settings (i.e., small sample sizes). I will show how domain knowledge can help guide architecture choices and efficiently make use of available data. In summary, there’s a critical need for machine learning in healthcare; however, the safe and meaningful adoption of these techniques requires close collaboration in interdisciplinary teams.
BIO: Jenna Wiens is a Morris Wellman Assistant Professor of Computer Science and Engineering (CSE) at the University of Michigan in Ann Arbor. Her primary research interests lie at the intersection of machine learning, data mining, and healthcare. Dr. Wiens received her PhD from MIT in 2014, was named Forbes 30 under 30 in Science and Healthcare in 2015, received an NSF CAREER Award in 2016, and was recently named to the MIT Tech Review's list of Innovators Under 35.
ABSTRACT: Data and analytics have always played a central role in the insurance industry. With the rapid development in digitalization and automation, insurance carriers have access to more of it than ever before. At State Farm, we intelligently integrate advanced analytics into every business process and take full responsibility of such technologies to help our customers’ life go right. This talk will discuss the opportunities for Machine Learning in an insurance setting, showcase some of the creative data science work that enables modernizing and transforming our organization, and also shed lights on the challenges of ethically and responsibly owning AI/ML solutions as an enterprise.
ABSTRACT: In the last 5 years, there have been a number of landmark research breakthroughs which pushed the fields of natural language processing, understanding, and generation forward in a significant way. Several advances have been made in natural language processing, understanding and generation (NLP, NLU, NLG), including transfer learning, more sophisticated language models, and novel approaches to content understanding. There are hundreds of relevant papers (and many more springing up every day) in NLP, NLU, and NLG. This new plethora of methods are having an impact for industry applications, and insurance is no exception. In this talk I will share several projects we are working on in American Family insurance where some of these recent NLP techniques are the key for improved performance and are the foundation to new and exciting applications in the insurance domain.
ABSTRACT: Many domains deal with large networks of interacting entities, such as social networks, the human brain, gene regulatory networks, and financial markets. For such networks that have emerged naturally, inferring the network topology, e.g. which entities interact with which other entities and how strongly, can be challenging. We will discuss several approaches for identifying the network topology or sparse approximations of the topology and related issues.
ABSTRACT: The spread of fake news was one of the most discussed characteristics of the 2016 U.S. Presidential Election. The concerns regarding fake news have garnered significant attention in both media and policy circles, with some journalists even going as far as claiming that results of the 2016 election were a consequence of the spread of fake news. Yet, little is known about the prevalence and focus of such content, how its prevalence changed over time, and how this prevalence related to important election dynamics. In this talk, I will address these questions by examining social media, news media, and interview data. These datasets allow examining the interplay between news media production and consumption, social media behavior, and the information the electorate retained about the presidential candidates leading up to the election.
ABSTRACT: HERE is the global leader in mapping and location technology. In this talk we acquaint the audience with HERE Technologies as a company and our HD Live map for autonomous driving. We then talk about various efforts within HERE Research to extract even more location information from the street-level imagery collected for the HD map.
Coffee, Bars, Yogurt, Fruit, etc.
ABSTRACT: Applied work in reinforcement learning has focussed on simulated environments that allow an agent to gather arbitrarily large quantities of data. Despite their successes, when data must be gathered in real time, algorithms in current use can require an eternity to gain competence, even in very simple environments. One critical issue is how an agent explores. Common approaches to exploration are highly inefficient, and I will discuss how this can be addressed through uncertainty representation and judicious probing. I will close by mentioning important directions for future work on exploration, learning from rich observations, and hierarchical representations, each of which may be essential to making reinforcement learning data-efficient.
BIO: Benjamin Van Roy is a Professor at Stanford University, where he has served on the faculty since 1998. His research focusses on understanding how an agent interacting with a poorly understood environment can learn over time to make effective decisions. Beyond academia, he leads a DeepMind Research team in Mountain View. He is a Fellow of INFORMS and IEEE, and in addition to those communities, he is a regular participant in ICML, NeurIPS, and RLDM. He has served on the editorial boards of Machine Learning, Mathematics of Operations Research, for which he co-edits the Learning Theory Area, Operations Research, for which he edited the Financial Engineering Area, and the INFORMS Journal on Optimization.
ABSTRACT: Conventional wisdom in machine learning taboos training on the test set, interpolating the training data, and optimizing to high precision. This talk will present evidence demonstrating that this conventional wisdom is wrong. I will additionally highlight commonly overlooked phenomena imperil the reliability of current learning systems: surprising sensitivity to how data is generated and significant diminishing returns in model accuracies given increased compute resources. I will close with a discussion of how new best practices to mitigate these effects are critical for truly robust and reliable machine learning.
BIO: Benjamin Recht is an Associate Professor in the Department of Electrical Engineering and Computer Sciences at the University of California, Berkeley. Prior to his time at Berkeley, he spent four magical years on the faculty of the CS department and WID at UW Madison. Ben's research group currently studies the theory and practice of optimization algorithms with a particular focus on applications in machine learning and control. Ben is the recipient of a Presidential Early Career Award for Scientists and Engineers, an Alfred P. Sloan Research Fellowship, the 2012 SIAM/MOS Lagrange Prize in Continuous Optimization, the 2014 Jamon Prize, the 2015 William O. Baker Award for Initiatives in Research, and the 2017 NeurIPS Test of Time Award.
ABSTRACT: As modern data science raises the demand for more efficient algorithms, a natural approach is to consider parallelizing computation. While such an approach has proven fruitful for many classical combinatorial algorithms, its usefulness in the domain of convex optimization – a workhorse of machine learning – is much less understood. In general, parallelization of a single iteration of convex optimization algorithms, such as in computing the gradient or Hessian of the objective function at a single point, is useful and has been exploited in practice. However, a roadblock to further speedups is reducing the overall number of iterations. Given the existing query lower bounds, such a speedup could only be achieved by querying many points in parallel per iteration (i.e., by using parallelization in the exploration of the feasible space). We show that it is generally not possible to reduce the iteration count of convex optimization algorithms by querying polynomially-many points per iteration, for essentially any interesting geometry (ell_p and Schatten_p spaces) and any level of the objective function smoothness. Most of the obtained lower bounds match the sequential complexity of these problems, up to, at most, a logarithmic factor in the dimension, and are, thus, (nearly) tight. Based on joint work with Cristóbal Guzmán. To appear in COLT 2019.
ABSTRACT: Given labeled points in a high-dimensional vector space, we seek a low-dimensional subspace such that projecting onto this subspace maintains some prescribed distance between points of differing labels. Intended applications include compressive classification. This talk will introduce a semidefinite relaxation of this problem, along with various performance guarantees. (Joint work with Culver McWhirter (OSU) and Soledad Villar (NYU).)
ABSTRACT: Mapping the human brain, or understanding how certain brain regions relate to specific aspects of cognition, has been and remains an active area of neuroscience research. Functional magnetic resonance imaging (fMRI) data—in the form of images, time series or graphs—are central in this research, but pose many challenges in the context of cognitive phenotype prediction as they are noisy and high dimensional with most available datasets providing relatively few training samples. Standardly employed linear models and newly proposed neural network methods pose limitations in complexity and interpretability, respectively, in this context. In this talk, I will present a new, interpretable neural network-based method, iBrainNN, which introduces the idea of node grouping into the design of the neural network. iBraiNN classifies cognitive performance from noisy fRMI-derived brain networks with 85% fewer model parameters than baseline deep models, and it is 2.6−69×faster while also identifying the most predictive brain sub-networks within several task-specific contexts.
BIO: Danai Koutra is an Assistant Professor in Computer Science and Engineering at University of Michigan, where she leads the Graph Exploration and Mining at Scale (GEMS) Lab. Her research focuses on practical and scalable methods for large-scale real networks, and has applications in neuroscience, organizational analytics, and social sciences. She won an NSF CAREER award and an Amazon Research Faculty Award in 2019, an ARO Young Investigator award and an Adobe Data Science Research Faculty Award in 2018, the 2016 ACM SIGKDD Dissertation award, and an honorable mention for the SCS Doctoral Dissertation Award (CMU). She is the Program Director of the SIAG on Data Mining and Analytics, an Associate Editor of ACM TKDD, a tutorial co-chair for KDD'19, and a demo co-chair for CIKM'19. At the University of Michigan, she is leading the "Explore Graduate Studies in CSE" workshop, which aims to broaden participation in computer science at the graduate level. She has co-organized 3 tutorials and 3 workshops. She has worked at IBM Hawthorne, Microsoft Research Redmond, and Technicolor Palo Alto/Los Altos. She earned her Ph.D. and M.S. in Computer Science from CMU in 2015 and her diploma in Electrical and Computer Engineering at the National Technical University of Athens in 2010.
ABSTRACT: With recent advances in machine learning, large enterprises incorporate machine learning models across a number of products. To facilitate training of these models, enterprises use shared, multi-tenant cluster of machines equipped with accelerators like GPUs. Similar to data analytics clusters, operators aim to achieve high resource utilization while providing resource isolation and fair sharing across users. In this talk we will first present characterization of machine learning workloads from a multi-tenant GPU cluster at Microsoft. We then study how various aspects of these workloads such as gang scheduling and locality constraints affect resource utilization and efficiency. Based on this analysis we discuss some research directions to improve efficiency and utilization both for individual jobs and across the cluster.
ABSTRACT: From routine movie recommendations to high-stakes medical treatment selection, machine learning is increasingly employed to augment or replace human decision makers. As a result, there is a growing need for predictive models which can accurately reflect complex decision-making goals. In predictive machine learning, the learning goal is most often formalized as a metric. This metric can then be used to evaluate and compare models, or a model can be trained to optimize the chosen metric directly. Importantly, optimizing the wrong metric can have an undesirable impact on the trained model behavior. This raises the question: given a learning problem, which metric should be used, so that the chosen metric best reflects the decision-maker preferences? This is the metric selection problem. One promising solution is metric elicitation; a framework for automatic metric selection via interactive user feedback. The metric is selected as one that best reflects implicit preferences. One key insight that enables efficient metric elicitation is that while machine learning models are complicated functions, metrics are most often designed to capture tradeoffs between low dimensional statistical quantities. For example, tradeoffs for many medical prediction problems can be reduced to determining the appropriate balance between false positive and false negative rates. Based on this property, I will present provably fast techniques for eliciting classification metrics using only pairwise preference queries.
BIO: Sanmi (Oluwasanmi) Koyejo an Assistant Professor in the Department of Computer Science at the University of Illinois at Urbana-Champaign. Koyejo's research interests are in the development and analysis of probabilistic and statistical machine learning techniques motivated by, and applied to various modern big data problems. He is particularly interested in the analysis of large scale neuroimaging data. Koyejo completed his Ph.D in Electrical Engineering at the University of Texas at Austin advised by Joydeep Ghosh, and completed postdoctoral research at Stanford University with a focus on developing Machine learning techniques for neuroimaging data. His postdoctoral research was primarily with Russell A. Poldrack and Pradeep Ravikumar. Koyejo has been the recipient of several awards including the outstanding NCE/ECE student award, a best student paper award from the conference on uncertainty in artificial intelligence (UAI) and a trainee award from the Organization for Human Brain Mapping (OHBM).
ABSTRACT: Today’s smart devices and AI computers run on semiconductor microchips. But creating chips to meet hyper-aggressive performance and cost targets requires exceptional semiconductor design and manufacturing control. Measuring chip features and finding defects near the atomic scale requires bleeding-edge lasers, optics, and precision motion control hardware. But these only produce raw data. The more critical technology is the data science transforming raw signals into process control insights, and AI is today’s breakthrough enabling tomorrow’s data science. AI of today is used to create the extreme computing which will drive the AI of the future.
We invite submissions of one page summary on all topics related to machine learning for the poster session to be held during the symposium.
Eligibility: We welcome abstracts from all members of machine learning community. There is no restriction on the eligibility for the poster presentation. However, only submissions with *first* author as an undergraduate or a graduate student will be considered for the Student Poster Award described below.
Submission guidelines: Submissions can be at most one page long, not including references. Do not include any supplementary files along with your submission. Posters based on papers previously published, or accepted for publication, etc. are permitted. The authors can add optional links to full papers/working papers/extended abstracts available in public domain such as arXiv, or open-source proceedings. However, please ensure that the abstracts are stand alone, as reviewers will not be required to review extra material.
Submission site: closed
Deadline for submission: May 6, 2019
Notification of accepted posters: May 16, 2019
Notification of finalists for student poster award: May 20, 2019
All accepted posters will be presented at MMLS June 6-7, 2019.
Poster size: The poster board is 4x4' and the poster should fit within, e.g., 48x36 inches.
Student Poster Award
We also have a Student Poster Award with rewards. There will be $1000 poster award for a maximum of three posters. All submissions with first authors as an undergraduate or a graduate student are eligible and will be automatically considered for the selection process. The selection process is as follows: A small number of eligible submissions with student authors will be selected as finalists. The finalists will be invited to give a short (3-5 min) spotlight presentation on the work. Judges will take into consideration the abstract, the poster, the spotlight presentation, as well as the presentation during the poster session in deciding the winners.
Award Candidates and Other Accepted Posters:
Check the link here for the list of posters.
The poster board is 4x4' and the poster should fit within, e.g., 48x36 inches. There will be numbers on the poster boards, and please post your poster on the board with the corresponding number. All posters will be presented in all poster sessions. You can put up your poster at 2:50pm on Thursday (before the first poster session), and take it down at 1:30pm on Friday (at the end of the last poster session).