Artificial Intelligence Seminar

Fall 2010
Friday 12:00-1:15
Upson 5130

The AI seminar will meet weekly for lectures by graduate students, faculty, and researchers emphasizing work-in-progress and recent results in AI research. Lunch will be served starting at noon, with the talks running between 12:15 and 1:15. The new format is designed to allow AI chit-chat before the talks begin. Also, we're trying to make some of the presentations less formal so that students and faculty will feel comfortable using the seminar to give presentations about work in progress or practice talks for conferences.

Date

Title/Speaker/Abstract/Host

September 3rd

Speaker: Chris J. Burges, Microsoft

Host: Thorsten Joachims

Title :  Ranking Research

Abstract :  Algorithms that learn to rank a large corpus of documents, given a query, form a core methodology in today's major Search Engines.  I will describe the approach that won Track 1 of Yahoo!'s recent Learning to Rank Challenge (ranking for Web search).  Our system is based on a general approach for learning arbitrary information retrieval measures.  I'll then briefly touch on some desiderata for our near term work on learning to rank, which suggest new research directions.  Finally I will describe a process that the Machine Learning Group at MSR Redmond is currently exploring in an attempt to identify long term, collaborative research ventures.

“The AI-Seminar is sponsored by Yahoo!”

September 10th

Speaker: Doug Turnbull , Assistant Professor at Ithaca College

Host: Thorsten Joachims

Bio : Doug Turnbull is currently a new assistant professor in the Department of  Computer Science at Ithaca College.  His main research interests are multimedia information retrieval, computer audition, machine learning, and human computation.

Doug received a B.S.E. degree (with honors) in Computer Science from Princeton University in 2001, and M.S. and Ph.D. degrees in Computer Science & Engineering from UC San Diego in 2005 and 2008.  While at UCSD, he co-founded the interdisciplinary Computer Audition Laboratory (CALab).  He spent the last two years at Swarthmore College where he was a visiting assistant professor.

Title :  Semantic Music Discovery Engine

Abstract :  Most commercial music discovery engines (including Apple iTunes Genius and Last.fm) rely on the analysis of social information (e.g., user preferences, blogs, or social tagging data, etc.) to help people find music. These systems are not ideal in that they suffer from popularity bias and the “cold start” problem.  To remedy these problems, researchers have been exploring content-based audio analysis as an alternative. However, state-of-the-art content-based systems often produce less-than-accurate annotations of music. It seems natural that we can improve music discovery by combining social and acoustic sources of music information.

My research focuses on developing a semantic music discovery engine: a system that an individual can use to discover music by describing what he or she wants to hear.  For example, you may want to find "funky bluegrass music with a prominent use of steel-string guitar".  In this talk, I will first describe how we can collect social information and use it to index music with semantically meaningful tags like "funky," "bluegrass," and "steel-string guitar.”  I will then describe a computer audition system that can automatically annotate music with tags based on analysis of the audio content.  Next, I will present a handful of algorithms that can be used to combine multiple sources of social and acoustic information.

Finally, I will show a demo of Meerkat , a web-based semantic music discovery engine, that is similar to a personalized Internet radio player like Pandora, but allows a user to control the stream of music using tags.  Meerkat was originally conceived as a collaborative class project in my undergraduate course on information retrieval. My students plan to launch a public version with Creative Commons music in the coming months.

“The AI-Seminar is sponsored by Yahoo!”

September 17th

Speaker: Ashutosh Saxena, Cornell University

Title :  "Make3D: Single Image Depth Perception and its applications to Robotics"

Abstract :  In this talk, I will talk about some of my recent learning algorithms that enable a robot to perceive its environment.

In particular, we will first consider the problem of converting standard digital pictures into 3D models. This is a challenging problem, since an image is formed by a projection of the 3D scene onto two dimensions, thus losing the depth information. We take a supervised learning approach to this problem, and model the scene depth as a function of the image features. We show that, even on unstructured scenes of a large variety of environments, our algorithm is frequently able to recover accurate 3D models. (See http://make3d.cs.cornell.edu ) We then look at the problem of combining our learning algorithm of single image depth estimation with other related sub-tasks in scene understanding (such as scene categorization, object detection). These sub-tasks operate on the same raw data and provide correlated outputs.

The last few decades have seen great progress in tackling each of these problems in isolation. Only, recently have researchers returned to the difficult task of considering them jointly. We consider learning a set of related models in such that they both solve their own problem and help each other. Our method requires only a limited “black box” interface with the models, allowing us to use very sophisticated, state-of-the-art classifiers without having to look under the hood.

We then apply our methods to robotics applications: (a) obstacle avoidance for autonomously driving a small electric car at high speeds through cluttered environments, and (b) robot manipulation, where we develop learning algorithms for grasping novel objects. This enables our robot to perform tasks such as open new doors, clear up cluttered tables, and unload items from a dishwasher.

“The AI-Seminar is sponsored by Yahoo!”

September 24th

Speaker: Alyosha Efros, CMU

Title :  Are Categories Necessary for Recognition?

Abstract :  The use of categories to represent concepts (e.g. visual objects) is so prevalent in computer vision and machine learning that most researchers don't give it a second thought. Faced with a new task, one simply carves up the solution space into classes (e.g. cars, people, buildings), assigns class labels to training examples and applies one of the many popular classifiers to arrive at a solution. In this talk, I will discuss a different way of thinking about object recognition -- not as object naming, but rather as object association. Instead than asking "What is it?", a better question might be "What is it like?"[M. Bar]. The etymology of the very word "re-cognize" (to know again) supports the view that association plays a key role in recognition. Under this model, when faced with a novel object, the task is to associate it with the most similar objects in one's memory which can then be used directly for knowledge transfer, bypassing the categorization step all-together. I will present some very preliminary results on our new model, termed "The Visual Memex", which aims to use object associations (in terms of visual similarity and spatial context) to reason about and parse visual scenes. We show that our model offers better performance at certain tasks than standard category-driven approaches.

Joint work with Tomasz Malisiewicz.

“The AI-Seminar is sponsored by Yahoo!”

October 1st

Speaker: Ruslan Salakhutdinov, MIT

Host: Ping Li

Bio: Ruslan Salakhutdinov received his PhD from University of Toronto, and is now a postdoctoral fellow at CSAIL and the department of Brain and Cognitive Sciences at MIT. His broad research interests involve developing flexible large-scale probabilistic models that contain deep hierarchical structure. Much of his current research concentrates on learning expressive structured representations using hierarchical models with applications to transfer learning. His other interests include Bayesian inference, matrix factorization, and approximate inference and learning of large-scale graphical models.

Title :  Learning Probabilistic Models with Deep Hierarchical Structures

Abstract :  Building intelligent systems that are capable of extracting higher-order knowledge from high-dimensional data and successfully transferring that knowledge to learning new concepts lies at the core of solving many AI related tasks, including object recognition, speech perception, and language understanding. Theoretical and biological arguments strongly suggest that building such systems requires models with deep architectures that involve many layers of nonlinear processing. In this talk I will first introduce a broad class of probabilistic generative models called Deep Boltzmann Machines (DBMs) that contain many layers of latent variables. I will describe a new learning algorithm for this class of models that uses variational methods and Markov chain Monte Carlo (MCMC). This new learning algorithm, in addition to a bottom-up pass, can incorporate top-down feedback, which allows DBMs to better propagate uncertainty about ambiguous inputs. I will further show that these deep models can learn interesting representations and can be successfully applied in many application domains, including information retrieval, object recognition, and nonlinear dimensionality reduction. In the second part of the talk, I will describe new ways of developing more complex systems that combine Deep Boltzmann Machines with more structured hierarchical Bayesian models. I will show how these hybrid models can learn a deep hierarchical structure for sharing knowledge across hundreds of visual categories, which allows efficient learning of new categories from few, even just one, examples -- a problem known as 'one-shot learning'.

“The AI-Seminar is sponsored by Yahoo!”

October 8th

Speaker: Jieping Ye, Arizona State University

Host: Ping Li

Date Change :*** WEDNESDAY October 6, 2010, 400pm - 5:00pm:

Location: G01 Biotechnology Building

Bio : Jieping Ye is an Associate Professor of the Department of Computer Science and Engineering at Arizona State University. He received his Ph.D. in Computer Science from University of Minnesota, Twin Cities in 2005. His research interests include machine learning, data mining, and biomedical informatics. He won the outstanding student paper award at ICML in 2004, the SCI Young Investigator of the Year Award at ASU in 2007, the SCI Researcher of the Year Award at ASU in 2009, the NSF CAREER Award in 2010, and the KDD best research paper award honorable mention in 2010.

Title :  Large-Scale Structured Sparse Learning

Abstract :  Recent advances in high-throughput technologies have unleashed a torrent of data with a large number of dimensions. Examples include gene expression pattern images, microarray gene expression data, and neuroimages. Variable selection is crucial for the analysis of these data. In this talk, we consider the structured sparse learning for variable selection where the structure over the features can be represented as a hierarchical tree, an undirected graph, or a collection of disjoint or overlapping groups. We show that the proximal operator associated with these structures can be computed efficiently, thus accelerated gradient techniques can be applied to scale structured sparse learning to large-size problems. We demonstrate the efficiency and effectiveness of the presented algorithms using synthetic and real data.

“The AI-Seminar is sponsored by Yahoo!”

October 15th

Speaker: Berkant Savas, University of Texas

Host: Charles Van Loan

Title :  Tools for large scale social network and graph computations

Abstract :  In this talk we will present a few novel tools for the link prediction and group recommendation problems in social network analysis. In particular, our discussion will contain three specific topics. (1) We will describe a technique called clustered low rank approximation, that captures and maintains fundamental structure from the social network. The main advantage of this method is significantly better low rank approximations that are obtained in less (or equal) amount of computation time. The improvements in the low rank approximation translate to improvements in the main task at hand, which include link prediction or group recommendation. The memory usage in the clustered approach is the same as for corresponding standard low rank approximations. This procedure is an effective and highly scalable tool for various tasks in computational analysis of social networks. (2) Suppose we are given an affiliation network between a set of users and a set of communities or groups, i.e. certain users are connected to certain groups. The task is then to give group recommendations to users that they may be interested in joining. Often, in addition to the affiliation network between the users and communities, there is a separate social network between the users themselves. We will discuss a few methods on how to combine the social network with the affiliation network in order to improve the performance for group recommendation. (3) Finally, we will discuss supervised methods for link predictions that utilize multiple and heterogeneous sources of information. We will show experimental results with real-world and large scale data sets on all three topics in the discussion.

"The AI-Seminar is sponsored by Yahoo!"

October 22nd

Speaker: Pannaga Shivaswamy, Post Doc

Host: Thorsten Joachims

Title :  Large Relative Margin and Applications

Abstract :  Over the last decade or so, machine learning algorithms such as
support vector machines, boosting etc. have become extremely popular.
The core idea in these and other related algorithms is the notion of
large margin . Simply put, the idea is to geometrically separate two
classes with a large separation between them; such a separator is then
used to predict the class of an unseen test example. These methods have
been extremely successful in practice and have formed a significant
portion of machine learning literature. There are several theoretical
results which motivate such algorithms. A closer look at such
theoretical results reveals that the generalization ability of these
methods are strongly linked to the margin as well as some measure of the
 spread of the data. Yet the algorithms themselves only seem to be
maximizing the margin ---completely ignoring the spread information. This
talk focuses on addressing this problem; novel formulations, that not
only take into consideration the margin but also the spread aspect of
the data, are proposed. In particular, relative margin machine, which is
 a strict generalization of the well known support vector machine is
proposed. Further, generalization bounds are derived for the relative
margin machines using a novel  method of landmark examples. The idea of
relative margin is fairly general; its potential is demonstrated by
proposing formulations for structured prediction problems as well as for
 a transductive setup using graph Laplacian. Finally, a boosting
algorithm incorporating both the margin information and the spread
information is derived as well. The boosting algorithm is motivated from
 the recent empirical Bernstein bounds.  All the proposed variants of
the relative margin algorithms are easy to implement, efficiently
solvable and typically show significant improvements over their large
margin counterparts--on real-world datasets.
(joint work with Dr. Tony Jebara)

“The AI-Seminar is sponsored by Yahoo!”

October 29th

Speaker: Surya Singh, University of Sydney

Host: Ashutosh Saxena

Bio: Surya Singh is a Research Fellow at the Australian Centre for Field Robotics where he leads modeling and control efforts. His research interests in dynamic systems include: agile motion over terrain, motion analysis, and mechatronics design. Dr. Singh is a Fulbright Scholar and has studied at the Stanford University's Robotic Locomotion Lab, Tokyo Institute of Technology's Hirose Robotics Lab, and Carnegie Mellon University's Field Robotics Center.

Title :  Agile Robots: Learning to Steer Naturally

Abstract :  Human locomotion can be remarkably informative for robotics. When it comes to agility, it can not only guide the mechanics (the ``how''), but also the steering (the ``where''). While this seems deceivingly simple -- one just accelerates, decelerates, or turns -- for dynamic systems it is complicated by inertia, saturation, compliance, and even social bias (e.g., driving direction). Determining the motion involves solving a complex inverse control problem. Analytic solutions are difficult to characterize. An adaptive method is introduced that uses human navigation to inform the planning and subsequent control. This is then illustrated in the context of robots -- with applications from controlling mining robots to observing animals in the field.

“The AI-Seminar is sponsored by Yahoo!”

November 5th

Speaker:

Host:

Bio :

Title

Abstract

“The AI-Seminar is sponsored by Yahoo!”

November 12th

Speaker: Katharina Morik, Technical University Dortmund

Host: Thorsten Joachims

Title :  Data Mining – Learning under Resource Constraints

Abstract :  Data Mining started in the nineties with the claim that real-world data collections as they are stored in data bases require less sophisticated and more scalable algorithms than the then dominating statistical routines. New tasks like frequent set mining occurred. At the same time, sophisticated pre-processing and sampling methods allowed data analysis to cope with large data sets.

Currently, we are again challenged by data masses at an even larger scale, collected at distributed sites, in heterogeneous formats and by applications that demand real-time response. Storage, runtime, and execution time for real-time behavior are the constrained resources, which need to be handled by new learning methods.

The talk will give an overview of learning under resource constraints and present applications that illustrate the new challenge, in more detail.

  • The overwhelming dimensionality of genomic data (about 200.000 features) demands fast and robust methods of stable feature selection . The small set of observations (about 100 patients) demands the integration of different populations. The two problems need to be solved, if we aim at a personalized medicine.
  • The new challenge is well illustrated by data analysis for ubiquitous systems. Logged data from a mobile device can be compressed by a data streaming algorithm such that further learning uses only the aggregated data. The prediction of file access allows decreasing upload-time and tailoring the operating system's services. In sum, this could save energy and let the battery last longer.

Implementing algorithms on GPGPUs is shortly discussed.

“The AI-Seminar is sponsored by Yahoo!”

November 19th

Speaker: NO SEMINAR - ACSU lunch

Host:

Bio :

Title

Abstract

“The AI-Seminar is sponsored by Yahoo!”

November 26th

Speaker: NO SEMINAR- Thanksgiving Break

Host:

Bio :

Title

Abstract

“The AI-Seminar is sponsored by Yahoo!”

December 3rd

 

 

Speaker:

Host:

Bio :

Title

Abstract

“The AI-Seminar is sponsored by Yahoo!”

December 10th

Speaker:

Host:

Bio :

Title

Abstract

“The AI-Seminar is sponsored by Yahoo!”

 

See also the AI graduate study brochure.

Please contact any of the faculty below if you'd like to give a talk this semester. We especially encourage graduate students to sign up! 

Sponsored by


CS7790, Fall '09
Claire Cardie
Carla Gomes
Joe Halpern
Dan Huttenlocher
Thorsten Joachims
Lillian Lee
Ashutosh Saxena
Bart Selman
Ramin Zabih

Back to CS course websites