Artificial Intelligence Seminar

Fall 2012
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.

August 20th

Speaker: Prof. Serge J. Belongie

Host:Theodoros Damoulas

Time: *** 2:00pm*** in 5130 Upson Hall

Title: Visual Recognition with Humans in the Loop

Abstract: We present an interactive, hybrid human-computer method for object classification. The method applies to classes of problems that are difficult for most people, but are recognizable by people with the appropriate expertise (e.g., animal species or airplane model recognition). The classification method can be seen as a visual version of the 20 questions game, where questions based on simple visual attributes are posed interactively. The goal is to identify the true class while minimizing the number of questions asked, using the visual content of the image. Incorporating user input drives up recognition accuracy to levels that are good enough for practical applications; at the same time, computer vision reduces the amount of human interaction required. The resulting hybrid system is able to handle difficult, large multi-class problems with tightly-related categories. We introduce a general framework for incorporating almost any off-the-shelf multi-class object recognition algorithm into the visual 20 questions game, and provide methodologies to account for imperfect user responses and unreliable computer vision algorithms. We evaluate the accuracy and computational properties of different computer vision algorithms and the effects of noisy user responses on a dataset of 200 bird species and on the Animals With Attributes dataset. Our results demonstrate the effectiveness and practicality of the hybrid human-computer classification paradigm.

“The AI-Seminar is sponsored by Yahoo!”

August 24th

Speakers: Ping Li, Cornell University

Title: One Permutation Hashing for Efficient Near Neighbor Search and Statistical Learning in BigData

Abstract: This work is an example that using basic statistics and probability in the right way may be able to substantially reduce the computational cost and energy-consumption for important industrial applications. The original minwise hashing algorithm is a standard technique widely deployed in the search industry; one typical application is to find near duplicates of Web pages.  Recently, we developed b-bit minwise hashing (Research Highlights in Comm. of the ACM 2011) by focusing only on a small number of bits of the hashed data and successfully applied b-bit hashing to: (1) training logistic regression and linear SVM on massive, extremely high-dimensional data (NIPS2011), and (2) fast near neighbor search by directly using the bits to construct hash tables (ECML2012). The major remaining problem is the preprocessing cost, as (b-bit) minwise hashing requires applying roughly 500 permutations on the entire data. This expensive preprocessing step could seriously affect the testing speed (on unprocessed examples) and cause considerable energy-consumptions. In this talk, our most recent (unpublished) work will demonstrate that merely ONE permutation is needed. Interestingly, one permutation hashing is even slightly more accurate (at 1/500 of the original cost).  We expect that this one permutation scheme (or its variants) will be adopted in practice. Joint work with Art Owen (Stanford Statistics) and Cun-Hui Zhang (Rutgers Statistics).

“The AI-Seminar is sponsored by Yahoo!”

August 31st

Speaker: Nir Ailon, Technion

Host: Thorsten Joachims

Title: Active Learning for Ranking and Clustering from Pairwise Information

Abstract: In this talk I will discuss two learning problems:  “Learning to Rank from Pairwise Preferences” and “Clustering from Pairwise Similarity Information”.  For both problems, traditional (passive) learning bounds are suboptimal.  In addition, general purpose active learning algorithms based on the disagreement coefficient are also suboptimal.  I will present a method for obtaining near optimal query complexity bounds for the two.  The method, called “Smooth Relative Regret Approximation” is an iterative algorithm relying on the ability, given a current hypothesis H, to build an empirical process approximating the difference between the loss of any hypothesis H’ and H, to within an error gracefully degrading as a function of the disagreement distance between H and H’.  Based on joint work with Ron Begleiter and Esther Ezra.

“The AI-Seminar is sponsored by Yahoo!”

September 7th

Speaker: Marcelo Finger

Host: Bart Selman

Bio: Marcelo Finger is a professor of Computer Science at the Department of Computer Science, University of Sao Paulo, Brazil, and is current a visiting academic in Cornell.  He obtained his MSc  (1990) and Phd (1994) from the Imperial College of Science and Technology of the University of London. His research interests include Artificial Intelligence, Computational Logics, Deductive and Probabilistic Reasoning.

Title : Probabilistic Satisfiability: Algorithms and Phase transition

Abstract: In this talk, we motivate the problem and present algorithms for probabilistic satisfiability (PSAT), an NP-complete problem, focusing on the presence and absence
of a phase transition phenomenon for each algorithm.  Our study starts by defining a PSAT normal form, on which all algorithms are based.  Several forms of reductions of PSAT to classical
propositional satisfiability (SAT) are proposed.  Theoretical and practical limitations of each algorithm are discussed.  Some algorithms are shown to present  a phase transition behavior. We show that variations of these algorithms may lead to the partial occlusion of the phase transition phenomenon and discuss the reasons for this change of practical behavior.

“The AI-Seminar is sponsored by Yahoo!”

September 14th

Speaker: Antonio Bahamonde

Host: Thorsten Joachims

Bio: Antonio Bahamonde is a Full Professor at the Department of Computer Sciences, University of Oviedo, Spain, and is current a visiting academic in Cornell. His research field is Machine Learning, both theoretical and practical applications.

Title: Beyond Classification

Abstract: The aim of the talk is to show some recent results that have in common the use of some extensions of binary or multi-class classification. Typically, these results use SVM or Logistic Regression. Included in this framework are applications to beef cattle selection, sensory studies of food products and bio-medical applications. Other interesting kind of extensions are provided by classifiers that predict a set of classes (usually called labels in this context) instead of single one; the so-called multilabel classifiers. The talk will include some work-in-progress ideas in this topic.

“The AI-Seminar is sponsored by Yahoo!”

September 21st

Speaker: Shuo Chen, Cornell University

Host:Thorsten Joachims

Title: Playlist Prediction via Metric Embedding

Abstract: Digital storage of personal music collections and cloud-based music services (e.g. Pandora, Spotify) have fundamentally changed how music is consumed. In particular, automatically generated playlists have become an important mode of accessing large music collections. The key goal of automated playlist generation is to provide the user with a coherent listening experience. In this paper, we present Logistic Markov Embedding (LME), a machine learning algorithm for generating such playlists. In analogy to matrix factorization methods for collaborative filtering, the algorithm does not require songs to be described by features a priori, but it learns a representation from example playlists. We formulate this problem as a regularized maximum-likelihood embedding of Markov chains in Euclidian space, and show how the resulting optimization problem can be solved efficiently. An empirical evaluation shows that the LME is substantially more accurate than adaptations of smoothed n-gram models commonly used in natural language processing.

Speaker: Karthik Raman, Cornell University

Title: Online Learning to Diversify from Implicit Feedback

Abstract: In order to minimize redundancy and optimize coverage of multiple user interests, search engines and recommender systems aim to diversify their set of results. To date, these diversification mechanisms are largely hand-coded or relied on expensive training data provided by experts. To overcome this problem, we propose an online learning model and algorithms for learning diversified recommendations and retrieval functions from implicit feedback. In our model, the learning algorithm presents a ranking to the user at each step, and uses the set of documents from the presented ranking, which the user reads, as feedback. Even for imperfect and noisy feedback, we show that the algorithms admit theoretical guarantees for maximizing any submodular utility measure under approximately rational user behavior. In addition to the theoretical results, we find that the algorithm learns quickly, accurately, and robustly in empirical evaluations on two datasets.

This is joint work with Thorsten Joachims and Pannaga Shivaswamy.

“The AI-Seminar is sponsored by Yahoo!”

September 28th

Speaker: Igor Labutov & Ian Lenz, Cornell University

Igor's Title: Humor as Circuits in Semantic Networks

Igor's Abstract: This work presents a first step to a general implementation of the Semantic-Script Theory
of Humor (SSTH). Of the scarce amount of research in computational humor, no research had focused on humor generation beyond simple puns and punning riddles. We propose an algorithm for mining simple humorous scripts from a semantic network (Concept- Net) by specifically searching for dual scripts that jointly maximize overlap and incongruity metrics in line with Raskin's Semantic-Script Theory of Humor. Initial results show that a more relaxed constraint of this form is capable of generating humor of deeper semantic content than wordplay riddles. We evaluate the said metrics through a user-assessed quality of the generated two-liners.

Ian's Title: Low-Power Parallel Algorithms for Single Image based Obstacle Avoidance in Aerial Robots

Ian's Abstract: For an aerial robot, perceiving and avoiding obsta- cles are necessary skills to function autonomously in a cluttered unknown environment. In this work, we use a single image captured from the onboard camera as input, produce obstacle classifications, and use them to select an evasive maneuver. We present a Markov Random Field based approach that models the obstacles as a function of visual features and non-local dependencies in neighboring regions of the image. We perform efficient inference using new low-power parallel neuromorphic hardware, where belief propagation updates are done using leaky integrate and fire neurons in parallel, while consuming less than 1 W of power. In outdoor robotic experiments, our algorithm was able to consistently produce clean, accurate obstacle maps which allowed our robot to avoid a wide variety of obstacles, including trees, poles and fences.

“The AI-Seminar is sponsored by Yahoo!”

October 5th

Speaker: NO SEMINAR- Fall Break


October 12th

Speaker: Jordan Pollack, Brandeis University

Host: Hod Lipson

Bio: Jordan Pollack is Professor and chair of computer science at Brandeis
University.. He received the Ph.D. from University of
Illinois in 1987. His lab, the Dynamical and Evolutionary Machine Organization (DEMO) has contributed to a broad range of different fields related to AI, like neural networks, dynamical systems, evolutionary computation, machine learning, cognitive science, artificial life, robotics, and educational technology.

Title : Towards Robot Embryogenesis

Abstract: In Nature, the embryogenesis process proceeds from a single fertilized
cell through division, migration, specialization and apoptosis.
Although a lot is known about development, we still have a long way
to go from theories of pattern formation towards understanding the
natural intelligence which robustly assembles complex biological
forms. And we are nowhere near robotic cells which can undergo mitosis.

Our approach has been to co-evolve bodies and brains in simulation
and then convert them into reality. I will review several generations of robots
which were automatically designed using co-evolutionary techniques.
The first generation was based on genetic programming and a simulation
of LEGO rod adhesion. The second generation, in collaboration with
Hod Lipson, used direct evolution on
an iterative simulation of truss structures, then used 3D printing for
the output. A third generation was based on generative representations
using L-systems driving a virtual factory, with manually assembly of
primitive servo modules. The current generation uses gene regulatory
networks to control an embryogenetic process where virtual cells
migrate, divide, and specialize into a small set of robotic components
which could be assembled using off the shelf technology.

“The AI-Seminar is sponsored by Yahoo!”

October 19th

Speaker: Hema S. Koppula, PhD Student, Cornell University

Host: Ashutosh Saxena

Title: Scene Understanding from RGBD data

Abstract: "Understanding human environments and human activities are two very important skills, especially for personal robots which operate in human environments. In this talk, I will present our work addressing these two tasks. We first consider the problem of semantic labeling of RGBD images of human environments. We propose a graphical model that captures various features and contextual relations, including the local visual appearance and shape cues, object co-occurence relationships and geometric relationships. In order to understand human activities, we consider the problem of extracting a descriptive labeling of the sequence of sub-activities being performed by a human, and more importantly, of their interactions with the objects in the form of associated affordances. Given a RGB-D video, we jointly model the human activities and object affordances as a Markov Random Field where the nodes represent objects and sub-activities, and the edges represent the relationships between object affordances, their relations with sub-activities, and their evolution over time. Our models admit efficient approximate inference, and we train it using a maximum-margin learning approach. We demonstrate the use of semantic labeling and descriptive labeling of activities in performing assistive tasks by a PR2 robot."

“The AI-Seminar is sponsored by Yahoo!”

October 26th

Speaker: Fei Sha

Host: Ping Li

Bio: Fei Sha is an assistant professor at the University of Southern California, Dept. of Computer Science. His primary research interests are machine learning and application to speech and language
> processing, computer vision, robotics and others. After obtaining his PhD under the supervision of Prof. Lawrence K. Saul from U. of Pennsylvania,  he worked as a postdoc with Profs. Michael I. Jordan and Stuart Russell at U. of California (Berkeley). He wrote his doctoral thesis on large margin based parameter estimation techniques for hidden Markov models. He has also worked
> extensively in dimensionality reduction. He has won outstanding student paper awards at NIPS and ICML.  He is a member of DARPA 2010 Computer Science Study Panel, and won an Army Research Office Young Investigator Award (2012).

Title: Domain Adaptation for Learning in a Changing Environment

Abstract: Statistical machine learning has become an important driving force behind many  application fields.  By large, however, its theoretical underpinning has hinged on the stringent assumption that the learning environment is stationary. In particular, the data distribution on which statistical models are optimized is the same as the distribution to which the models are applied. Real-world applications are far more complex than the pristine condition. For instance, computer vision systems for recognizing objects in images often suffer from significant performance degradation if they are evaluated on image datasets that are different from the dataset on which they are designed.

In this talk, I will describe our efforts in addressing this important challenge of building intelligent systems that are robust to distribution disparity. The central theme is to learn invariant features and adapt probabilistic models across different distributions (i.e., domains). To this end, our key insight is to discover and exploit hidden structures in the data. These structures, such as manifolds and discriminative clusters, are intrinsic and thus resilient to distribution changes due to exogenous factors. I will present several learning algorithms we have proposed and demonstrate their effectiveness in pattern recognition tasks from computer vision and natural language processing.

This talk is based on the joint work with my students (Boqing Gong and Yuan Shi, both from USC) and our collaborator Prof. Kristen Grauman (U. of Texas, Austin).

“The AI-Seminar is sponsored by Yahoo!”

November 2nd

Speaker: **Cancelled** Phil Long, NEC Labs

Host: Ping Li

Title: On the Necessity of Irrelevant Variables

Abstract: Abstract: An irrelevant variable typically decreases the accuracy of a classifier; after all, it makes the predictions of the classifier depend to a greater extent on random chance. We show, however, that the harm from irrelevant variables can be much less than the benefit from relevant variables, so that it is possible to learn very accurate classifiers, almost all of whose variables are irrelevant.  It can be advantageous to continue adding variables, even as their prospects for being relevant fade away.  We showed this with theoretical analysis and experiments using artificially generated data (so that we would know which variables were relevant and irrelevant). Both of these use an assumption, conditional independence, formalizing the intuitive idea that variables are not redundant.
In the situation that we studied relatively few of the many variables are relevant, and the relevant variables are only weakly predictive. In this case, algorithms that cast a wide net outperform more selective algorithms whose hypotheses, on average, contain more relevant than irrelevant variables.

(This is joint work with Dave Helmbold of UC Santa Cruz.)

“The AI-Seminar is sponsored by Yahoo!”

November 9th

Speaker: "Cancelled"





“The AI-Seminar is sponsored by Yahoo!”

November 16th Speaker: NO SEMINAR- ACSU Lunch

November 23rd


November 30th

Speaker: Abhinav Gupta, Carnegie Mellon University

Host: Ashutosh Saxena

Bio: Abhinav Gupta is an Assistant Research Professor at the Robotics Institute, Carnegie Mellon University. Prior to this, he was a postdoctoral fellow at CMU working with Alexei Efros and Martial Hebert. His research is in the area of computer vision, and its applications to robotics and computer graphics. He is particularly interested in using physical, functional and causal relationships for understanding images and videos. His other research interests include exploiting relationship between language and vision, semantic image parsing, and exemplar-based models for recognition. Abhinav received his PhD in 2009 from the University of Maryland under Prof. Larry Davis. His dissertation was nominated for the ACM Doctoral Dissertation Award by the University of Maryland. Abhinav is a recipient of the ECCV Best Paper Runner-up Award (2010) and the University of Maryland Dean’s Fellowship Award (2004).

Title: Beyond Naming: Image Understanding via Rich Representations

Abstract: What does it mean to "understand" an image? One popular answer is simply naming the objects seen in the image. During the last decade most computer vision researchers have focused on this "object naming" problem. While there has been great progress in detecting things like "cars" and "people", such a level of understanding still cannot answer even basic questions about an image such as "What is the geometric structure of the scene?", "Where in the image can I walk?" or "What is going to happen next?". In this talk, I will present three different type of representations which help us to develop deeper understanding of the visual world: (1) Firstly, I will talk about physically and geometrically based representations that are meaningfully grounded in the real world. (2) Next, I will introduce human-centric representation where we represent and reason about space from the point of view of a human agent. (3) Finally, I will briefly discuss representations where understanding is itself formulated as an association problem.

“The AI-Seminar is sponsored by Yahoo!”

December 7th






“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, Spring '12
Claire Cardie
Carla Gomes
Joe Halpern
Dan Huttenlocher
Thorsten Joachims
Lillian Lee
Ashutosh Saxena
Bart Selman
Ramin Zabih

Back to CS course websites