[cs-talks] Upcoming CS Seminars: NRG (Mon) + PhD Proposal (Tues) + IVC Seminar (Tues)
cs at bu.edu
Fri Feb 6 11:50:15 EST 2015
Revisiting Network Resource Allocation in Data Centers
Fahad Dogar, Tufts
Monday, February 9, 2015 at 11:00am in MCS 148
Abstract: Popular applications like Facebook and Google Search perform rich and complex "tasks" (e.g., generating a user's social newsfeed). From a network perspective, these tasks typically comprise multiple flows, which traverse different parts of the network at potentially different times. Existing network resource allocation schemes (e.g., TCP), however, treat all these flows in isolation - rather than as part of a task - which delays completion of tasks (i.e., user requests). In this talk, I will make a case for "task-aware" network scheduling, and present Baraat, a decentralized task-aware scheduling system. Compared to existing approaches (e.g., TCP and other flow based schemes), Baraat improves both the average and tail response times for a wide range of workloads. I will also present a deployment friendly transport framework (PASE) which can support richer resource allocation schemes (e.g., task-aware scheduling) without requiring changes to network switches. Both Baraat and PASE appeared in ACM Sigcomm 2014. Joint work with researchers from MSR, MSU, and LUMS.
Bio: Fahad Dogar is an assistant professor in the computer science department at Tufts University. Earlier he did his PhD from Carnegie Mellon and undergrad from LUMS, Pakistan. Most recently, he was a post-doc in the systems and networking group at Microsoft Research UK. Webpage:
>From Object Semantics to Large-scale Summarization
Dr. Leonid Sigal, Disney Research
Tuesday, February 10, 2015 at 2pm in MCS B29
Abstract: In the first part of the talk I will describe a discriminative semantic model for object categorization. The model is based on semantic manifold embedding, where we embed both images, class prototypes and auxiliary semantic entities such as supercategories and attributes. By exploiting such a unified model for semantics, we ensure that each category can be represented by the corresponding supercategory + a sparse combination of attributes, with an additional exclusive regularization to learn discriminative composition. This model also generates compact semantic description of each category, which enhances generalization, interoperability and enables humans to analyze what has been learned. In the second part of the talk I will focus on the framework for large scale image set and video summarization. Starting from the intuition that the characteristics of the two media types are different but complementary, we develop a fast and easily-parallelizable approach for creating not only video summaries but also novel structural summaries of events in the form of the storyline graphs. The storyline graphs can illustrate various events or activities associated with the topic in the form of a branching directed network. The video summarization is achieved by diversity ranking on the similarity graphs between images and video frame, thereby treating consumer image as essentially a form of weak-supervision. The reconstruction of storyline graphs on the other hand is formulated as inference of the sparse time-varying directed graphs from a set of photo streams with assistance of consumer videos. Time permitting I will also talk about a few other recent project highlights.
Bio: Leonid Sigal is a Senior Research Scientist at Disney Research Pittsburgh and an adjunct faculty at Carnegie Mellon University. Prior to this he was a postdoctoral fellow in the Department of Computer Science at University of Toronto. He completed his Ph.D. at Brown University in 2008; he received his B.Sc. degrees in Computer Science and Mathematics from Boston University (1999), his M.A. from Boston University (1999), and his M.S. from Brown University (2003). From 1999 to 2001, he worked as a senior vision engineer at Cognex Corporation, where he developed industrial vision applications for pattern analysis and verification. Leonid's research interests mainly lie in the areas of computer vision, machine learning, and computer graphics. He has published more than 50 papers in venues and journals in computer vision, computer graphics and machine learning (including publications in PAMI, IJCV, CVPR, ICCV, ECCV, NIPS, UAI, and ACM SIGGRAPH). His current research spans articulated pose estimation, action recognition, object detection and categorization, transfer learning, latent variable models, data-driven simulation, controller design for animated characters and perception of human motion.
Learning Space-Time Structures for Human Action Recognition and Localization
Shugao Ma, BU
Tuesday, February 10, 2015 at 3:30pm in MCS 148
Abstract: In this thesis we study the problem of action recognition and localization in realistic video clips. In training only the action class labels of the training video clips are available, and in testing both the action label and the spatial location of the action performer(s) are to be predicted. Although many past works have been done, this remains a challenging problem due to the complexity of the human actions and the large intra-class variations. Human actions are inherently structured patterns of body movements. However, past works are inadequate in learning the space-time structures in human actions and leveraging them for action recognition. In this thesis we propose new methods that exploit such space-time structures for effective human action recognition and localization. In the feasibility study, we developed a new local space-time representation for action recognition and localization, the hierarchical Space-Time Segments . Using this new representation, we explored ensembles of hierarchical spatio-temporal trees, discovered directly from training data, to model these structures for action recognition. This proposed approach in the feasibility study achieved state-of-the-art performances on two challenging benchmark datasets UCF-Sports and HighFive. However, further works are needed to more efficiently discover space-time structures and better handle large scale data. In the remaining work, we will explore deep convolutional neural network (CNN) for larger scale action recognition problem, studying ensemble of CNN models trained on whole video frames blended with motion information and CNN models trained on automatically proposed foreground regions. We will also explore sub-graph vectorization method that can effectively encode space-time structures of human actions into vectors, which will enable us to efficiently discover discriminative structures in a vector space. We will evaluate the remaining works on larger scale dataset, e.g. , the UCF101 dataset that has 101 action classes and 13000 videos.
Do-Not-Track and the Economics of Third-Party Advertising
Georgios Zervas, BU
Junior Faculty Fellow, Hariri Institute for Computing
Assistant Professor of Marketing, School of Management
Wednesday, February 11, 2015 at 3pm in MCS 180 - Hariri
Abstract: Retailers regularly target users with online ads based on their web browsing activity, benefiting both the retailers, who can better reach potential customers, and content providers, who can increase ad revenue by displaying more effective ads. The effectiveness of such ads relies on third-party brokers that maintain detailed user information, prompting legislation such as do-not-track that would limit or ban the practice. We gauge the economic costs of such privacy policies by analyzing the anonymized web browsing histories of 14 million individuals. We find that only 3% of retail sessions are currently initiated by ads capable of incorporating third-party information, a number that holds across market segments, for online-only retailers, and under permissive click-attribution assumptions. Third-party capable advertising is shown by 12% of content providers, accounting for 32% of their page views; this reliance is concentrated in online publishing (e.g., news outlets) where the rate is 91%. We estimate that most of the top 10,000 content providers could generate comparable revenue by switching to a “freemium” model, in which loyal site visitors are charged $2 (or less) per month. We conclude that do-not-track legislation would impact, but not fundamentally fracture, the Internet economy.
Bio: Georgios Zervas is an Assistant Professor of Marketing at Boston University’s School of Management. Before joining BU in 2013 he was a Simons postdoctoral fellow at Yale and an affiliate at the Center for Research on Computation and Society at Harvard. He received his PhD in Computer Science in 2011 from Boston University. He is broadly interested in understanding the strategic interactions of firms and consumers participating in internet markets using large-scale data collection and econometric analysis.
Computational Understanding of Image Memorability
Zoya Bylinskii, MIT
Thursday, February 19, 2015 at 4pm in MCS 148
Abstract: In this talk, I will describe the research done in the Oliva Lab on Image Memorability - a quantifiable property of images that can be used to predict whether an image will be remembered or forgotten. Apart from presenting the lab's research directions and findings, I will focus on the work I have done in understanding and modeling the intrinsic and extrinsic factors that affect image memorability. I will present results on how consistent people are in which images they find memorable and forgettable (across experiments, settings, and visual stimuli) and I will show how these findings generalize to information visualizations. I will also demonstrate how the extrinsic factors of image context and observer eye behavior modulate image memorability. I will present an information-theoretic model of context and image distinctiveness, to quantify their effects on memorability. Finally, I will demonstrate how eye movements, pupil dilations, and blinks can be predictive of image memorability. In particular, our computational model can use an observer's eye movements on an image to predict whether or not the image will be later remembered. In this talk, I hope to offer a more complete picture of image memorability, including the contributions to cognitive science, and the computational applications made possible.
The following is the first paper on image memorability that has come out of the Oliva Lab, and has started a whole direction of research:http://cvcl.mit.edu/papers/IsolaXiaoTorralbaOliva-PredictingImageMemory-CVPR2011.pdf -- it can give people some background, though I will provide an intro as well.
Bio: Zoya Bylinskii is a PhD student at MIT, jointly supervised by Aude Oliva and Fredo Durand. She works in the area of computational perception - at the intersection of cognitive science and computer science. Specifically, she is interested in studying human memory and attention, in order to build computational models to advance the understanding and application possibilities of these areas. Her current work spans a number of research directions, including: image memorability, saliency benchmarking, and information visualizations. Zoya most recently completed her MS under the supervision of Antonio Torralba and Aude Oliva, on a "Computational Understanding of Image Memorability". Prior to this, her BS research on parts-based object recognition was supervised by Sven Dickinson at the University of Toronto. She also spent a lovely summer in 2011 working in BU with Stan Sclaroff on reduplication detection in sign language :)
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cs-talks