Artificial Intelligence Seminar

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 4	Art Munson, Cornell University Host: Thorsten Joachims On Feature Selection, Bias-Variance, and Bagging This will be a practice talk for a presentation Art is making at the European Conference on Machine Learning next week. The talk will be understandable by a general audience. Abstract: We examine the mechanism by which feature selection improves the accuracy of supervised learning. An empirical bias/variance analysis as feature selection progresses indicates that the most accurate feature set corresponds to the best bias-variance trade-off point for the learning algorithm. Often, this is *not* the point separating relevant from irrelevant features, but where increasing variance outweighs the gains from adding more (weakly) relevant features. In other words, feature selection can be viewed as a variance reduction method that trades off the benefits of decreased variance (from the reduction in dimensionality) with the harm of increased bias (from eliminating some of the relevant features). If a variance reduction method like bagging is used, more (weakly) relevant features can be exploited and the most accurate feature set is usually larger. In many cases, the best performance is obtained by using all available features. This is a joint work with Rich Caruana.
September 11	*CANCELLED* Haym Hirsh, NSF and Rutgers Host: Jon Kleinberg
September 18	Ping Li, Cornell Host: Ashutosh Saxena Recent Development in Predictive Learning and Boosting: ABC-Boost, Robust LogitBoost and ABC-LogitBoost Abstract: Friedman's MART (multiple additive regression trees) algorithm is widely used in industry for classification and ranking (e.g., search engines). This talk mainly presents ABC-LogitBoost, which considerably outperforms ABC-MART (Ping Li, 2009), MART (Friedman, 2001), and LogitBoost (Friedman et al, 2000) for multi-class classification, on a wide variety of large-scale datasets. The talk will also present Robust LogitBoost, which fixed a known numerical problem in the original LogitBoost and outperformed MART and other classification algorithms for binary classification on many datasets. ABC-Boost stands for the general framework of "adaptive base class boost" and both ABC-LogitBoost and ABC-MART are specific implementations of ABC-Boost.
September 25	Carla Brodley, Tufts Host: Claire Cardie No AI-Seminar at the regular time, but talk in the afternoon at MLDG.
October 2	Ashutosh Saxena, Cornell Robot Learning: Single Image Depth Perception and Robotic Grasping Abstract: The ability to perceive the 3D shape of the environment is a basic ability for a robot. We present an algorithm to convert 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. We then apply our methods to robotics applications: (a) obstacle avoidance for autonomously driving a small electric car at high speeds through cluttered environments, (b) autonomous helicopter flight in indoor constrained environments, and (c) 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.
October 9	* No Seminar *
October 16	Gaurav Pandey, University of Minnesota, Twin Cities Host: Ashutosh Saxena Data Mining Techniques for Enhancing Protein Function Prediction Predicting the cellular functions of proteins is one of the most important goals of computational biology, and data mining and machine learning techniques applied to genomic data are playing an important role in achieving this goal. This talk will cover our research on techniques that can utilize unutilized or under-utilized information in genomic data sets to improve the performance of traditional approaches for protein function prediction. In particular, we will focus on the problem of incorporating inter-relationships between functional classes, as captured in the Gene Ontology, into standard protein function prediction algorithms. To address this problem, we incorporate these inter-relationships into the k-nearest neighbor classifier, where the strength of the inter-relationships is computed using a standard measure for evaluating the semantic similarity between nodes in a hierarchy of labels, or an ontology. This incorporation improves the accuracy of the function predictions made, and also enables the discovery of interesting biology in the form of novel functional annotations for several yeast proteins, such as Sna4, Rtn1 and Lin1. Also, although our work is discussed in the context of protein function prediction, similar ideas and techniques are also expected to be useful for other problems involving multi-label classification, such as text and image classification. More details of our work can be found here.
October 23	Andy Ruina, Cornell Host: Ashutosh Saxena TBA
October 30	Regina Barzilay, MIT Host: Lillian Lee Learning to follow orders: Reinforcement learning for mapping instructions to actions In this talk, I will address the problem of relating linguistic analysis and control --- specifically, mapping natural language instructions to executable actions. This technique has enabled automation of tasks that until now have required human participation --- for example, automatically configuring software by consulting how-to guides. Our results demonstrate that this method can rival supervised techniques while requiring few or no annotated training examples. Joint work with Branavan, Harr Chen and Luke Zettlemoyer.
November 6	Hadas Kress-Gazit, Cornell University Host: Ashutosh Saxena Abstract: High-level tasks to correct, low-level robot control Robots today can mop the floor, assist surgeons and explore space; however, there is no robot that could be trusted to drive autonomously in a real city. Robots either perform simple or hard-coded tasks fully autonomously or they operate with close human supervision. While most of the sensing and actuation technology required for high-level operation exists, what is lacking is the ability to plan at a high-level while providing guarantees for safety and correctness of a robot's autonomous behavior. In this talk I will present a formal approach to creating robot controllers that ensure the robot satisfies a given high level task. I will describe a framework in which a user specifies a complex and reactive task in Structured English. This task is then automatically translated, using temporal logic and tools from the formal methods world, into a hybrid controller. This controller is guaranteed to control the robot such that its motion and actions satisfy the intended task, in a variety of different environments.
November 13	Theodoros Damoulas, Cornell University Host: Thorsten Joachims Probabilistic Multiple Kernel Learning   The integration of multiple and possibly heterogeneous information sources for an overall decision-making process has been an open and unresolved research direction in computing science since its very beginning. This talk will summarize research that addresses parts of this direction by proposing  probabilistic  data integration algorithms for  multiclass  decisions, where an observation of interest is assigned to one of many categories based on a  plurality  of information channels.   The adopted Bayesian probabilistic framework is motivated by the requirements for assessing decision-making costs, formal inclusion of prior knowledge and principled model selection. Requirements that are common across many fields, such as bioinformatics and robotics, where multiple sources of information are available for a multiclass classification decision.
November 20	Rob Fergus, NYU Host: Dan Huttenlocher ****** TIME CHANGE: 1:00PM - 2:15PM ****** ROOM CHANGE: 315 Upson Hall ****ALADDINS Serving lunch Title: Semi-Supervised Learning in Gigantic Image Collections. Abstract:    With the advent of the Internet it is now possible to collect    hundreds of millions of images. These images come with varying    degrees of label information. “Clean labels” can be manually    obtained on a small fraction, “noisy labels” may be extracted    automatically from surrounding text, while for most images there are    no labels at all. Semi-supervised learning is a principled framework    for combining these different label sources. However, it scales    polynomially with the number of images, making it impractical for    use on gigantic collections with hundreds of    millions of images and thousands of classes. In this paper we show    how to utilize recent results in machine learning to obtain highly    efficient approximations for semi-supervised learning that are    linear in the number of images. Specifically, we use the convergence    of the eigenvectors of the normalized graph Laplacian to    eigenfunctions of weighted Laplace-Beltrami operators. Our algorithm    enables us to apply semi-supervised learning to a database of 80    million images gathered from the Internet.    Joint work with Yair Weiss (Hebrew U.) and Antonio Torralba (MIT).    Short Bio:    Rob Fergus is an assistant professor of computer science at the    Courant Institute, New York University. He received his PhD from    University of Oxford, jointly advised by Prof. Andrew Zisserman and    Prof. Pietro Perona at Caltech. Before coming to NYU he was a    postdoc at MIT with Prof. William T. Freeman.
November 27	*No Seminar*
December 4	*No Seminar*

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

Back to CS course websites