We are pleased to announce the new Yahoo seminar series in Machine Learning.
There is currently a wide range of research in the area of machine
learning at Rutgers; the goal of this seminar series is to spread
awareness of research in this topic inside the university and across
disciplines where it is studied and leveraged. It is hoped we may
have speakers from many different departments including but not
limited to computer science, psychology, statistics, economics as well
as speakers from organizations outside the university, such as Yahoo.
Meetings are held Tuesdays at 11:00 am at CORE A (301), and notification will be given by email as to the exact schedule of the various topics covered.
The colloquia is graciously sponsored by Yahoo, and is organized by Pavel Kuksa, Michael Littman, Vladimir Pavlovic, and Ari Weinstein
Link to the 2010/2011 Series
Previously Held Talks:
2010
- February:
- 2:
Ying Hung, webpage
Title: Analysis of Computer Experiments with Functional Response
Computer experiments refer to those experiments that are performed
in computers using physical models and finite element analysis. Most
existing methods for analyzing computer experiments with single
outputs such as kriging cannot be easily extended to functional
outputs due to the computational problems caused by high-
dimensionality of the response. In this talk, we develop an
efficient implementation of kriging for analyzing functional
responses. The methodology is illustrated using computer
experiments conducted for machining process optimization and data
center thermal management.
2009
- January:
- 26:
John Langford, webpage
Title: Algorithms for Learning with Partial Labels
Abstract: Learning can be thought of as competing with a large set of
functions. The foundations and theory for this are well understood in a
fully supervised setting where the value (or loss) of all answers is
known. In a more natural setting, only the chosen action's value is
revealed. Example of this more natural setting are endemic to
applications of machine learning, such as predicting the optimal search
result, ad, news story, etc... The foundations of learning in this
"partial label" setting are just now being worked out. I'll provide an
overview of results, as well as some new algorithms that advance the
frontier of what we understand how to do.
- December:
- 15:
Umar Syed, webpage
Title: Algorithms for Apprenticeship Learning
I will describe several algorithms for apprenticeship learning: the problem of learning to behave in an environment with an unknown reward function by observing the behavior of an expert. We follow on the work of Abbeel and Ng (2004), who considered a framework in which the unknown reward function is assumed to be a linear combination of a set of known and observable features. Our algorithms, like theirs, are guaranteed to learn a policy that is nearly as good as the expert?s, given enough examples. However, our algorithms have the following additional advantages: (1) Their running time has a much weaker dependence on the number of features, (2) they do not require an expensive post-processing step that invokes a quadratic program solver, (3) they can output stationary policies, (4) they sometimes produce a policy that is substantially better than the expert?s, and (5) they sometimes output policies which enjoy a certain "lexicographic" optimality property. I will also show a demonstration of the algorithms on a toy video-game environment.
Joint work with Rob Schapire, Michael Bowling
- 1:
Nicholas Belkin, webpage
Title: Personalizing Information Retrieval: Discovering Behavioral Correlates of Task Type, Topic Knowledge and Document Usefulness
Department of Library and Information Science, School of Communication & Information
Performance increases in information retrieval (IR), whether in Web search engines or digital
libraries (DLs), depends crucially on being able to tailor system interaction to characteristics of
the individual user and that person's context. Factors that have been shown to influence what a
person finds useful (or to predict what a user would find useful) include the nature of the task
that led the person to engage in information seeking, the stage of task accomplishment that the
person is in, the person's level of knowledge of the topic of interest, how long the person looks at
a document, and other features of previous and current behaviors. In this talk, I outline our
current research on the problem of Personalization of the Digital Library Experience
(http://comminfo.rutgers.edu/imls/poodle). The general goal of this project is to develop a
"personalization assistant", an application that sits on a person's computing devices, monitors the
person's behaviors and context, and uses this information to personalize the person's information
seeking interactions. We are currently analyzing data from several experiments in which we have
controlled the type of task that a person is working on, the stage the person is in the task
completion process, and the topic of the search, and have elicited such features as the level of
knowledge of the task and of the topic, and some of the person's cognitive abilities. The data that
we have collected has to do with a variety of search behaviors, including various features of the
queries that the person posed during the searching episode, mouse movements, scrolling, time
spent on web page, eye movements, and saving and using behaviors, as well as usefulness
judgments of web pages saved for the task. We have two goals for the data analysis at this point:
one is to discover interactions among the different personal and contextual factors in prediction
of usefulness of documents; the other is to discover relationships between our independent
variables (the personal and contextual factors, and the usefulness judgments), and observed
behaviors (the dependent variables). We have some ideas about how to accomplish these goals,
but are hoping to get more and perhaps better ideas through discussion at this seminar.
- November
- 17:
Stephen Jose Hanson, webpage
Title: Brain Reading: Multivariate Classification of Mental States
Psychology, Rutgers, Newark Campus
RUMBA Labs
Recent trends in brain imaging research have been focusing on what appears to be novel Multivariate statistical tools, that in point of fact are quite familiar to those in the neural computation and machine learning fields. Nonetheless, these older methods are critical to the challenges in this research include the BOLD signal which provides less than few per-cent of relevant signal variation, unknown multivariate structure, non-gaussian signal distributions, and high dimensional (50k-400k) data sets which are severely underdetermined by available training examples. I will review the motivation for these new trends, present new applications including the nature of the socalled "face area", generalization between individual brain states and explore the resultant decision surfaces and visualization of cortical similarity.
- 10:
Tong Zhang, webpage
Title: High Dimensional Nonlinear Learning using Local Coordinate Coding
Abstract:
We present a new method for learning nonlinear functions in high dimension using semisupervised learning. Our method includes a phase of unsupervised basis learning and a phase of supervised function learning. The learned bases provide a set of anchor points to form a local coordinate system, such that each data point on a high dimensional manifold can be locally approximated by a linear combination of its nearby anchor points, with the linear weights offering its local-coordinate coding. We show that a high dimensional nonlinear function can be approximated by a global linear function with respect to this coding scheme, and the approximation quality is ensured by the locality of such coding. The method turns a difficult nonlinear learning problem into a simple global linear learning problem, which overcomes some drawbacks of traditional local learning methods. The empirical success of our method has been demonstrated in a recent pascal image classification competition, where the top performance was achieved by an NEC system using this approach.
Joint work with Kai Yu at NEC Lab America.