Select a calendar:
Filter June Events by Event Type:
SUNMONTUEWEDTHUFRISAT
University Calendar
Events for June
-
Ph.D. Defense - Michael Tsang 6/11 2:00 pm
Thu, Jun 11, 2020 @ 02:00 PM - 04:00 PM
Thomas Lord Department of Computer Science
University Calendar
Ph.D. Defense - Michael Tsang 6/11 2:00 pm "Interpretable Machine Learning Models via Feature Interaction Discovery"
Ph.D. Candidate: Michael Tsang
Date: Thursday, June 11, 2020
Time: 2:00 PM - 4:00 PM
Committee: Yan Liu (Chair), Emily Putnam-Hornstein, Xiang Ren
Title: Interpretable Machine Learning Models via Feature Interaction Discovery
Abstract:
The impact of machine learning prediction models has created a growing need for us to understand why they make their predictions. The interpretation of these models is important to reveal their fundamental behavior, to obtain scientific insights into data, and to help us trust automatic predictions. In this dissertation, we examine how to explain black-box prediction models via feature interaction detection and attribution, i.e. if features influence each other and how these interactions contribute to predictions, respectively.
We first discuss how feature interaction detection leads to model interpretations of diverse domains such as image/text classification and automatic recommendation. Here, we focus on the special case of recommendation where interaction detection improves not only model interpretability but also prediction performance. We then discuss how to attribute predictions to feature interactions in a way that is simultaneously interpretable, model-agnostic, principled, and scalable. Our discussion culminates in the unification of interaction detection and attribution to yield general prediction visualizations that are both intuitive and insightful.
Meeting Links:
Zoom Meeting:
https://usc.zoom.us/j/5669704161
Meeting ID: 566 970 4161
Google Meet (ONLY A BACKUP - IF WE EXPERIENCE PROBLEMS WITH ZOOM):
https://meet.google.com/brt-fjya-ykd
Phone Number:
(US) +1 720-439-6997
PIN:455 863 061
WebCast Link: https://usc.zoom.us/j/5669704161
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
WEBINAR SERIES: Digital Technologies for COVID-19
Fri, Jun 12, 2020 @ 11:00 AM - 12:00 PM
USC Viterbi School of Engineering
University Calendar
This week our webinar will feature talks by Dean Yannis Yortsos, Vice Dean Assad Oberai, and Prof. Urbashi Mitra from the USC Viterbi School of Engineering. Please find more details of our speakers and their talks below.
Talk 1: A spatiotemporal SIR model for modeling the spread of an infectious disease - by Yannis Yortsos, Assad Oberai and Harisankar Ramaswamy
Abstract: A continuum model for the spread of an infectious disease is developed by drawing an analogy with the interaction of multiple reacting chemical species. The resulting set of partial differential equations is analyzed and insights are drawn by considering several limiting states. Finally, a simple finite-element scheme is developed to solve these equations and to make predictions corresponding to several what-if scenarios.
Bio: Dr. Yannis Yortsos is the Dean of the USC Viterbi School of Engineering and the Zohrab Kaprielian Chair in Engineering, a position he has held since 2005. He received a BS (Diploma) degree in Chemical Engineering from the National Technical University of Athens, Greece, and MS and PhD degrees from the California Institute of Technology -- all in chemical engineering. His research area is in fluid flow, transport and reaction processes in porous media with specific application to the subsurface. He was elected to the National Academy of Engineering in 2008, where he has also served as Secretary, Vice Chair and Chair of Section 11. Since July 2017, Dr. Yortsos has served as a member of the NAE Council. Along with colleagues at Duke University and Olin College, he co-founded in 2009 the Global Grand Challenges Scholars Program, now adopted by many universities in the U.S. and overseas. He organized and hosted at USC in Fall 2010 the NAE Second Grand Challenges Summit, which spurred in 2013 the Global Grand Challenges Summits. Between 2012 and 2017, Dr. Yortsos was the Chair of the Diversity Committee of the Engineering Deans Council, in which capacity he spearheaded an engineering diversity initiative, now adopted by more than 210 engineering deans nationwide. In recognition of these initiatives, the USC Viterbi School of Engineering received in 2017 the ASEE President's Award, and was one of the four engineering schools nationwide that received the ASEE Award for Excellence in Veterans in Engineering. Dr. Yortsos is the PI of the NSF I-Corps Innovation Node Los Angeles, established in 2014 as a partnership between USC, Caltech and UCLA.
Bio: Assad Oberai is Hughes Professor of Aerospace and Mechanical Engineering and Interim Vice Dean of Research at the USC Viterbi School of Engineering. Prior to joining USC, he was a Professor in the Department of Mechanical Aerospace and Nuclear Engineering at Rensselaer Polytechnic Institute (RPI), the Associate Dean for Research and Graduate Studies in the School of Engineering, and the Associate Director of the Scientific Computation Research Center (SCOREC). Assad was an Assistant Professor of Aerospace and Mechanical Engineering at Boston University from 2001 to 2005, and joined the Rensselaer faculty in 2006. He received a PhD from Stanford University in 1998, an MS from the University of Colorado in 1994, and a Bachelor's degree from Osmania University in 1992, all in Mechanical Engineering. Assad is on the Editorial Board of three journals, including PlosOne.
Talk 2: Group Testing for Efficient SARS-CoV-2 Assessment - by Urbashi Mitra
Abstract: Several challenges to wide-scale virus testing have emerged during the course of the pandemic within the United States. There have been insufficient tests, a lack of key testing elements as well as widely varying accuracies amongst the tests. We review group testing which was introduced in the 1940s by Dorfman, and, more generally, our recent work on active hypothesis testing. Our research focus has been on making decisions with maximized accuracy within a finite amount of time versus minimizing the average stopping time which has been the classical approach. Prior approaches have investigated the asymptotic accuracy of the strategies, whereas we can assess performance for a finite number of observations for our method. We compare and contrast other group testing methods that have been suggested by other research teams. Our numerical results show that one can effectively combine tests of differing accuracies with re-testing to strongly reduce the number of tests needed to test a population. Thoughts on how to extend these ideas to different kinds of testing goals (testing for the virus versus testing for antibodies) will also be provided.
Bio: Urbashi Mitra received the B.S. and the M.S. degrees from the University of California at Berkeley and her Ph.D. from Princeton University. Dr. Mitra is currently the Gordon S. Marshall Professor in Engineering at the University of Southern California. She was the inaugural Editor-in-Chief for the IEEE Transactions on Molecular, Biological and Multi-scale Communications. Dr. Mitra is a Fellow of the IEEE. She is the recipient of: the 2017 IEEE Women in Communications Engineering Technical Achievement Award, a 2015 UK Royal Academy of Engineering Distinguished Visiting Professorship, a 2015 US Fulbright Scholar Award, a 2015-2016 UK Leverhulme Trust Visiting Professorship, IEEE Communications Society Distinguished Lecturer, 2012 Globecom Signal Processing for Communications Symposium Best Paper Award, 2012 US National Academy of Engineering Lillian Gilbreth Lectureship, the 2009 DCOSS Applications & Systems Best Paper Award, Texas Instruments Visiting Professorship (Fall 2002, Rice University), 2001 Okawa Foundation Award, 2000 Ohio State University's College of Engineering Lumley Award for Research, 1997 Ohio State University's College of Engineering MacQuigg Award for Teaching, and a 1996 National Science Foundation CAREER Award. She has been an Associate Editor for multiple IEEE Transactions.
Dr. Mitra has held visiting appointments at: King's College, London, Imperial College, the Delft University of Technology, Stanford University, Rice University, and the Eurecom Institute. Her research interests are in: wireless communications, communication and sensor networks, biological communication systems, detection and estimation and the interface of communication, sensing and control.
Co-hosted by:
Craig Knoblock, Executive Director, USC Information Sciences Institute
Bhaskar Krishnamachari, Director, USC Viterbi Center for CPS and IoTWebCast Link: https://usc.zoom.us/webinar/register/WN_bS2IGZDMTw2aymiLaBAzIw
Audiences: Everyone Is Invited
Contact: Bhaskar Krishnamachari
-
PhD Defense - Liron Cohen
Wed, Jun 17, 2020 @ 10:00 AM - 12:00 PM
Thomas Lord Department of Computer Science
University Calendar
Ph.D. Defense - Liron Cohen 6/17 10:00am "Efficient Bounded-Suboptimal Multi-Agent Path Finding and Motion Planning via Improvements to Focal Search"
Ph.D. Candidate: Liron Cohen
Date: Wednesday, June 17, 2020
Time: 10:00 AM - 12:00 PM
Committee: Sven Koenig (Chair), Bistra Dilkina, Peng Shi, Maxim Likhachev
Location: Online due to COVID-19
Zoom: https://usc.zoom.us/j/4421939378
Google Meet (only if there are issues with Zoom): meet.google.com/ova-uozj-vej
Title: Efficient Bounded-Suboptimal Multi-Agent Path Finding and Motion Planning via Improvements to Focal Search
Abstract:
Cooperative autonomous agents navigating in different environments can be useful in real-world application domains such as warehouse automation, search-and-rescue, fighting forest fires, traffic control, computer games and mining. Agents are required to navigate in such complex environments while avoiding obstacles and each other. In Artificial Intelligence (AI), a simplified model of this problem is called the Multi-Agent Path Finding (MAPF) problem. In MAPF, time is discretized into timesteps and the environment is discretized into cells that individual agents can occupy exclusively at any given timestep. Along with the environment, the MAPF problem also specifies unique start and goal cells for each agent. A solution to the MAPF problem is a set of paths, one for each agent, that take the agents from their respective start cells to their respective goal cells without collisions. A path for an agent is a sequence of move (between adjacent cells) or wait (at a cell) actions, each with some duration in timesteps.
Different measures for the cost of a solution to the MAPF problem, such as sum-of-costs or makespan, are commonly used in AI. Unfortunately, finding an optimal solution to a MAPF problem according to any of these cost measures is NP-hard, thus explaining why optimal MAPF solvers are inefficient in many real-world application domains. On the other end of the spectrum, suboptimal MAPF solvers are efficient but can be ineffective or even incomplete. One framework that balances the trade-off between efficiency and effectiveness is that of bounded-suboptimality. A bounded-suboptimal solver takes a user-specified constant w>=1 and returns a solution with a cost guaranteed to be at most w times the cost of an optimal solution. On the one hand, the freedom to explore suboptimal solutions allows these solvers to be more efficient than optimal ones. On the other hand, and as opposed to suboptimal solvers, these solvers ensure that their solutions are effective.
Unfortunately, state-of-the-art bounded-suboptimal MAPF solvers, all of which rely heavily on a bounded-suboptimal heuristic search method called Focal Search (FS), have a few shortcomings. First, small changes in w can significantly affect these solvers' runtime. Thus, it is often difficult to determine a w such that an effective solution is found efficiently. Second, these solvers are inefficient when agents are huddled together and the environment has some structural components that create bottlenecks. One such example is a typical warehouse domain. Here, the environment has long narrow corridors connecting open spaces or other corridors, and agents have to move between different regions of the warehouse through the corridors. Third, these solvers are inefficient when move actions have different durations. One such example is actions that model kinodynamically feasible motions. These shortcomings make bounded-suboptimal MAPF solvers inefficient in many real-world application domains, thus making suboptimal MAPF solvers the only viable option. This inefficiency is problematic because a low solution cost is often important for performing the task at hand sufficiently well.
In this dissertation, we improve FS in ways that resolve the above shortcomings of bounded-suboptimal MAPF solvers. Specifically, we develop an anytime framework for FS, which alleviates the need to choose w carefully. An anytime bounded-suboptimal MAPF solver based on this framework finds a ``good'' solution quickly and refines it to better and better solutions if time allows. We also show that FS provides bounded-suboptimality guarantees even when it is used with inflated heuristics. We develop an inflated heuristic, called the highway heuristic, which improves the efficiency of bounded-suboptimal MAPF solvers when the domain's environment has structural components that create bottlenecks. The highway heuristic alters the paths that agents choose by biasing the agents away from their individual optimal paths and towards paths on shared ``highways'' (directional lanes) in the environment. This results in implicit coordination of the agents, which, in turn, increases the efficiency of bounded-suboptimal MAPF solvers. Finally, we develop a version of FS that reasons about wait durations ``in bulk'' by using timeintervals instead of timesteps. A bounded-suboptimal MAPF solver that uses this version is efficient even when move actions have different durations, thus making it suitable for solving MAPF problems with actions that model kinodynamically feasible motions.
On the theoretical side, we formally prove that our improvements are bounded-suboptimal. On the experimental side, we show that our improvements result in increased efficiency in domains inspired by real-world applications. The overall impact of this dissertation is twofold. The first impact is in opening up the possibility of using bounded-suboptimal MAPF solvers for new application domains. New application domains include ones in which existing bounded-suboptimal MAPF solvers are not applicable, such as when the agents are kinodynamically constrained (for example, forklifts), as well as ones in which existing bounded-suboptimal MAPF solvers are too inefficient, such as automated warehouses. Different benefits of using bounded-suboptimal MAPF solvers in such application domains include safety (that is, the solutions are guaranteed to be collision-free), completeness and effectiveness (since the cost of the solution is guaranteed to be at most w times the cost of an optimal solution). The second impact is in revitalizing FS. Compared to other heuristic search methods, such as A* and wA*, FS is less restricted in the states it can choose to expand and can take advantage of a broader family of heuristics. Nevertheless, A* has received significantly more attention than FS from the scientific community.\footnote{According to Google Scholar, as of May 15, 2020, the first paper to introduce A* has been cited 9,522 times while the first paper to introduce FS has been cited 83 times.} Our improvements to FS make it more applicable and call for revisiting its importance.
WebCast Link: https://usc.zoom.us/j/4421939378
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Max Pflueger
Wed, Jun 17, 2020 @ 04:00 PM - 05:30 PM
Thomas Lord Department of Computer Science
University Calendar
Ph.D. Defense - Max Pflueger 6/17 4:00 pm "Learning from Planners to Enable New Robot Capabilities"
Ph.D. Candidate: Max Pflueger
Date: Wednesday, June 17, 2020
Time: 4:00 pm - 5:30 pm
Committee: Gaurav S. Sukhatme (chair), Joseph Lim, Sandeep Gupta, Ali Agha
Title: Learning from Planners to EnableNew Robot Capabilities
Abstract:
Solving complex robotic problems often involves reasoning about behavior multiple steps in the future, yet many robot learning algorithms do not incorporate planning structures into their architecture. In this dissertation we show how we can harness the capabilities of planning algorithms to learn from the structure of the robotic problems we wish to solve, thus expanding beyond what was available from baseline planners. We consider problems in multi-arm manipulation planning, path planning for planetary rovers, and reinforcement learning for torque controlled robots, and show how in each case it is possible to learn from the behavior of planning algorithms that are limited and unable to solve the full generalized problem. Despite not being full solutions these planners provide useful tools and insights that can be leveraged in larger solutions.
In multi-step planning for manipulation we develop a high level planner that can find solutions in difficult spaces by solving sub-problems in sub-spaces of the main planning space. For planetary rovers show how to use inverse reinforcement learning to learn a new planning algorithm that can function on different (and generally cheaper) input data. Reinforcement learning algorithms often suffer from unstable or unreliable training, we show how this can be mitigated by augmenting the robot state with a state embedding space learned from planner demonstrations.
Planning and control algorithms often rely on rigid and prescribed assumptions about the nature of robot problems, which may not be suitable for the generalized and versatile robot systems we wish to build. However, as this dissertation argues, those structures are still useful in informing the behavior of more flexible families of algorithms.
Meeting Links:
Zoom Meeting:
https://usc.zoom.us/j/93816840971
Meeting ID: 938 1684 0971
Google Meet (ONLY A BACKUP - IF WE EXPERIENCE PROBLEMS WITH ZOOM):
Meeting ID:
https://meet.google.com/iuj-gygb-utw
WebCast Link: https://usc.zoom.us/j/93816840971
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
WEBINAR SERIES: Digital Technologies for COVID-19
Fri, Jun 19, 2020 @ 11:00 AM - 12:00 PM
USC Viterbi School of Engineering
University Calendar
This week our webinar will feature a talk by Prof. Dave Conti and Dr. Abigail Horn. Prof. Conti is in Division of Biostatistics in the Department of Preventive Medicine and the Norris Comprehensive Cancer Center (NCCC) at USC, and Dr. Horn is a Postdoctoral Fellow in the Department of Preventive Medicine at USC.
Incorporation of risk factors in a stochastic epidemiological COVID-19 model for Los Angeles County -- by Dave Conti & Abigail Horn
Abstract: We have developed an epidemic compartmental model to address the key questions of (1) When and how the epidemic dynamics will impact health care capacity? (2) What happens to the dynamics of the epidemic when social distancing changes? and (3) How will the epidemic affect different at-risk groups? Our model uses stochastic differential equations and approximate Bayes calculation techniques for parameter estimation. We incorporate external information to inform prior distributions for parameter specification. This includes previous studies on risk factors for COVID-19 to inform differences in illness severity (e.g. advanced age, existing health conditions) and the prevalence of these risk factors in Los Angeles County; mobility data to inform time-varying changes in contact rate; and (3) seroprevalence data to estimate the fraction of unobserved illnesses.
Our modeling framework enables modifying parameters at different time points, enabling the specification of interventions, e.g. social distancing scenarios. A key contribution of our model over existing models estimating hospital resource demand is that it accounts for combinations of risk factors (age, unhealthy behaviors, existing comorbidities, and combinations of comorbidities), including area-level differences in prevalence of these risk factors, in model dynamics.
Accounting for differential risk for specific populations allows us to estimate the impact of COVID-19 on these populations in Los Angeles, including analyzing findings by race/ethnicity groups, to inform the prioritization of these populations for protection. For example, this allows us to compare our projections to the observed disparities in death rates across race/ethnicity groups and answer the question of whether these differences can be explained by prevalence of risk factors alone, or if other factors (e.g. disparities in exposure, differences in contact rates) must be involved. We find that death rates are higher for these populations than can be explained by the risk factors alone.
Model-estimated parameters, projections, and probabilities of severe illness for different combinations of risk factors are provided online on our project website: uscbiostats.github.io/COVID19
Bios:
Dave Conti, PhD
Dr. Conti is Professor in the Division of Biostatistics in the Department of Preventive Medicine and the Norris Comprehensive Cancer Center (NCCC) at USC. He is Associate Director for Data Science Integration for the NCCC at USC and the Kenneth T. Norris, Jr. Chair in Cancer Prevention. His research focuses on study design and statistical methods for genetic and environmental epidemiology. His methodological research aims to integrate biological knowledge in statistical modeling and he has several past R01s to develop and investigate the use of Bayesian hierarchical models for genomic studies and to develop statistical methods to integrate genetic and omics data. He has developed statistical and bioinformatics software for analysis, most recently this includes LUCIDus (https://cran.r-project.org/web/packages/LUCIDus/index.html) and hJAM (https://cran.r-project.org/web/packages/hJAM/index.html).
Dr. Conti is currently Director of the Data Science Core for a U19 project investigating aggressive prostate cancer in African-American men integrating the built environment, germline and somatic genetic profiles, gene expression and tumor microenvironment data. He is also Co-Investigator on a P01 focusing on the development of statistical methods for integrated genomics analysis, as well as numerous applied epidemiology studies.
Abigail Horn, PhD
Dr. Horn is a Postdoctoral Fellow in the Department of Preventive Medicine at the University of Southern California and a member of the Center for Applied Network Analysis (CANA). Her research interests involve network epidemiology, probabilistic modeling and data science in the context of public health, with a focus on foodborne diseases and diseases of diet. She received a Ph.D. from the Institute for Data, Systems and Society at MIT and a Bachelor's in Physics from the College of Creative Studies at UCSB.
Prior to joining USC, she led a research project at the German federal-level food protection agency to develop, implement, and evaluate algorithms and decision support systems for modeling food supply networks to identify the source of large-scale outbreaks of foodborne disease. Her work has been funded by the NIH, the Robert Wood Johnson Foundation, the Bayer Foundation, the German Research Foundation, and the Santa Fe Institute.
Co-hosted by:
Craig Knoblock, Executive Director, USC Information Sciences Institute
Bhaskar Krishnamachari, Director, USC Viterbi Center for CPS and IoTWebCast Link: https://usc.zoom.us/webinar/register/WN_bS2IGZDMTw2aymiLaBAzIw
Audiences: Everyone Is Invited
Contact: Bhaskar Krishnamachari
-
COVID-19: Droplet and Airborne Transmission
Fri, Jun 26, 2020 @ 11:00 AM - 12:00 PM
Viterbi School of Engineering Alumni
University Calendar
Assessing the 6ft rule and occupation of enclosed spaces
A key societal need in the presence of SARS-CoV-2 is to safely occupy enclosed spaces for extended periods of time (e.g., offices, classrooms, community facilities etc.). This requires a characterization of the respiratory emissions from individuals with and without masks for a single breath or a single sneeze but must also examine the effect of multiple people in a confined space for durations measured in hours. Transmission of SARS-CoV-2 via larger respiratory droplets and direct contact has been established. Several studies suggest that airborne transmission may also be important. Fluid dynamics plays an essential role in determining transmission via both mechanisms. In this seminar, we will present preliminary results from a collaborative research effort that combines experiments, numerical simulations, and theoretical modeling in fluid dynamics to characterize aerosol and droplet dispersion in scenarios ranging from host-to-host, near-range transmission, to room-scale, long-duration exposure events that characterize different return-to-work settings.WebCast Link: https://usc.zoom.us/webinar/register/WN_bS2IGZDMTw2aymiLaBAzIw
Audiences: Everyone Is Invited
Contact: Kristy Ly
