Artificial Intelligence Seminar

February 10th

Speaker: David Smith, UMass

Host: Claire Cardie

Title: Inferring and Exploiting Relational Structure in Large Text Collections

Abstract: The digitization of knowledge and concerted retrospective scanning projects are making overwhelming amounts of text in diverse domains, genres, and languages available to readers and researchers. To make this data useful, our group is working on improving OCR, language modeling, syntactic analysis, information extraction, and information retrieval. I will focus in particular on problems of inferring the relational structure latent in large collections of documents, such as books, web pages, patent applications, grant proposals, and social media ostings. Which books or passages quote, translate, paraphrase, and cite each other? This research requires improvements in modeling translation and other forms of similarity, as well as improvements in efficiently comparing large numbers of passages. Finally, I will discuss how passage similarity relations can be used to improve tasks such as named-entity recognition and syntactic parsing.

“The AI-Seminar is sponsored by Yahoo!”

February 17th

Speaker: Stefanie Tellex

Host: Hadas Kress-Gazit

Bio: Stefanie Tellex is a Research Scientist at the MIT Computer Science and Artificial Intelligence Laboratory. A native of Rochester, she completed her Ph.D. at the MIT Media Lab in 2010, where she developed models for the meanings of spatial prepositions and motion verbs. She has published at SIGIR, HRI, AAAI, IROS, and ICMI, winning Best Student Paper at SIGIR and ICMI. Her research interests include probabilistic graphical models, human-robot interaction, and grounded language understanding.

Title: Understanding Language of Movement and Manipulation

Abstract: Natural language is a compelling modality for controlling complex systems such as robots, with its promise of powerful, intuitive interaction. However, robustly understanding language from untrained users is a challenging problem. In this talk I describe a probabilistic approach to understanding natural language commands
given to robots. The framework, called Generalized Grounding Graphs, defines a probabilistic graphical model that maps between constituents in the language and objects and actions in the external world. The framework learns models for the meanings of complex verbs such as "put" and "take," as well as spatial relations such as "on" and "to." The model allows efficient inference and learning by using the compositional structure of a natural language command to factor the distribution over interpretations. This factorization enables it to compose learned word meanings and understand novel commands that have never been previously encountered. The system is trained and evaluated using parallel corpora of language paired with robot actions collected using crowd sourcing. Grounding graphs are a first step towards robots that can robustly interact with a human partner using natural language.

“The AI-Seminar is sponsored by Yahoo!”

February 24th

Speaker: Cristian Danescu-Niculescu-Mizil, Cornell University

Host: Claire Cardie

Title : Language as Influence(d)

Abstract: What effect does language have on people, and what effect do people have on language?  The answers to these questions can help shape the future of social-media systems by bringing a new understanding of communication and collaboration between users.

I will describe two of my efforts to address these fundamental problems computationally, exploiting very large-scale textual and social data.  The first project uncovers previously unexamined contextual biases that people have when determining which opinions to focus on, using Amazon.comhelpfulness votes on reviews as a case study to evaluate competing theories from sociology and social psychology. The second project leverages insights from psycho- and socio-linguistics and embeds them into a novel computational framework in order to provide a new understanding of how key aspects of social relations between individuals are embedded in (and can be inferred from) their conversational behavior.  In particular, I will discuss how power differentials between interlocutors are subtly revealed by how much one individual immediately echoes the linguistic style of the person they are responding to.

This talk includes joint work with Susan Dumais, Michael Gamon, Jon Kleinberg, Gueorgi Kossinets, Lillian Lee and Bo Pang.

“The AI-Seminar is sponsored by Yahoo!”

March 2nd

Speaker:

Host:

Bio:

Title :

Abstract:

“The AI-Seminar is sponsored by Yahoo!”

March 9th

Speaker: Thomas Gärtner, Universitaet Bonn

Host: Thorsten Joachims

Title : Machine Learning for Dynamic Difficulty Adjustment in Computer Games

Abstract: While difficulty adjustment is common practise in many traditional games (consider, for instance, the handicaps in Golf and Go), the case for dynamic difficulty adjustment in electronic games has been made only recently. To date, most computer games only have static difficulty settings and computer game researchers have proposed a number of heuristic approaches. In this talk, I (i) formalise dynamic difficulty adjustment as a learning problem on partially ordered sets, (ii) propose an exponential update algorithm for this setting, (iii) show a bound on the number of wrong difficulty settings relative to the best static setting chosen in hindsight, and (iv) demonstrate the empirical performance of the algorithm.

“The AI-Seminar is sponsored by Yahoo!”

March 16th

Speaker: Ashish Raj

Host: Ramin Zabih

Title : Graph Theoretic Analysis of Human Brain Networks in Health and Disease

Abstract: Abstract: Whole brain connectivity networks were derived from various MR modalities: cortical thickness association networks from structural MRI, tract connectivity networks from diffusion MRI, and (future work) correlation networks from fMRI data. These networks allow us to interrogate various network-level features of both healthy and diseased brains. In this talk I will show that
a) cortical thickness networks can distinguish between healthy, severe and mild epileptic patients
b) connectivity networks from diffusion MRI reveal hierarchical but hub-free organization in the brain
c) the brain optimally places cortical regions in order to minimize wiring cost, and achieves optimal information flow at the cheapest cost
d) diffusion processes on brain networks can reproduce the spatial patterns of several well-known dementias.

“The AI-Seminar is sponsored by Yahoo!”

March 23rd

Speaker: NO SEMINAR- Spring Break

Host:

Bio:

Title :

Abstract

“The AI-Seminar is sponsored by Yahoo!”

March 30th

Speaker: NO SEMINAR- ACSU Lunch

Host:

Bio:

Title :

Abstract:

“The AI-Seminar is sponsored by Yahoo!”

April 6th

Speaker: Myle Ott, Cornell University

Title: Estimating the Prevalence of Deception in Online Review Communities (WWW 2012)

Abstract: Consumers' purchase decisions are increasingly influenced by user-generated online reviews. Accordingly, there has been growing concern about the potential for posting deceptive opinion spam---fictitious reviews that have been deliberately written to sound authentic, to deceive the reader. But while this practice has received considerable public attention and concern, relatively little is known about the actual prevalence, or rate, of deception in online review communities, and less still about the factors that influence it.
We propose a generative model of deception which, in conjunction with a deception classifier, we use to explore the prevalence of deception in six popular online review communities: Expedia, Hotels.com, Orbitz, Priceline, TripAdvisor, and Yelp. We additionally propose a theoretical model of online reviews based on economic signaling theory, in which consumer reviews diminish the inherent information asymmetry between consumers and producers, by acting as a signal to a product's true, unknown quality. We find that deceptive opinion spam is a growing problem overall, but with different growth rates across communities. These rates, we argue, are driven by the different signaling costs associated with deception for each review community, e.g., posting requirements. When measures are taken to increase signaling cost, e.g., filtering reviews written by first-time reviewers, deception prevalence is effectively reduced.

This talk includes joint work with Claire Cardie and Jeff Hancock.

Speaker 2: Ruben Sipos, Cornell University

Title: Large-Margin Learning of Submodular Summarization Models (EACL 2012)

Abstract: In this paper, we present a supervised learning approach to training submodular scoring functions for extractive multi-document summarization. By taking a structured prediction approach, we provide a large-margin method that directly optimizes a convex relaxation of the desired performance measure. The learning method applies to all submodular summarization methods, and we demonstrate its effectiveness for both pairwise as well as coverage-based scoring functions on multiple datasets. Compared to state-of-the-art functions that were tuned manually, our method significantly improves performance and enables high-fidelity models with number of parameters well beyond what could reasonably be tuned by hand.
This talk includes joint work with Pannaga Shivaswamy and Thorsten Joachims.

“The AI-Seminar is sponsored by Yahoo!”

April 13th

Speaker:

Host:

Bio:

Title :

Abstract:

“The AI-Seminar is sponsored by Yahoo!”

April 20th

Speaker: Haiyuan Yu, Cornell University

Host: Joe Halpern

Title : Understanding large-scale biological networks

Abstract: Almost all biological processes involve protein-protein interactions. The set of all protein interactions in the cell is referred to as its “interactome”, which is often modeled as a network (or graph). It has long been believed that decoding such networks is necessary for the understanding of human disease. We have developed a high-quality high-throughput pipeline to screen through tens of millions of protein pairs to generate a comprehensive interactome map for a given organism. We also established a statistical framework to help design such large-scale experiments and, more importantly, quantitatively and experimentally measure the quality of each detected interaction, as well as the whole set. With the publication of several large-scale protein-protein interactome networks, including ours, researchers have recently begun to use complex cellular networks to delineate underlying mechanisms for human disorders and predict unknown disease genes, but with little success. One main reason is that most analyses model proteins strictly as graph-theoretical nodes, ignoring structural details and spatial constraints of proteins and their interactions. We developed a “homology modeling” approach to determine at atomic resolution the interface of each interaction to generate the first three-dimensional (3D) interactome network for human disease. We then systematically examined relationships between 3949 genes, 62663 mutations and 3453 associated disorders within this 3D network, leading to the prediction of 292 candidate genes for 694 unknown disease-to-gene associations with concrete molecular mechanism hypotheses. Furthermore, proteins are often organized into modules through their interactions to carry out the same functions, with many proteins belonging to multiple modules. Traditional clustering algorithms are often wrongly applied to infer these functional modules. We developed a novel clustering algorithm to detect such overlapping modules within protein networks.

“The AI-Seminar is sponsored by Yahoo!”

April 27th

Speaker: Claudio Silva, NYU

Host: Carla Gomes

Bio: Claudio T. Silva is Professor of Computer Science and Engineering at the Polytechnic Institute of NYU. From 2003 to 2011, he was with the School of Computing and the Scientific Computing and Imaging Institute at the University of Utah. He coauthored more than 175 technical papers and eight U.S. patents, primarily in visualization, geometric processing, scientific data management, and related areas. He received IBM Faculty Awards in 2005, 2006, and 2007, and best paper awards at IEEE Visualization 2007, IEEE Shape Modeling International 2008, the 2010 Eurographics Educator Program, the ACM Eurographics Symposium on Parallel Graphics and Visualization 2011, and EuroVis 2011. His work is (or has been) funded by grants from the NSF, NIH, DOE, IBM, and ExxonMobil.

Title : Exploratory Visualization

Abstract: We take the view that future advances in science, engineering, and medicine depend on the ability to comprehend the vast amounts of data being produced and acquired. Visualization is a key enabling technology in this endeavor: it helps people explore and explain data through software systems that provide a static or interactive visual representation. Despite the promise that visualization can serve as an effective enabler of advances in other disciplines, the application of visualization technology is non-trivial. The design of effective visualizations is a complex process that requires understanding of existing techniques and how they relate to human cognition. For a visualization to be insightful, it needs to be both effective and efficient. This requires a combination of design and science to reveal information that is otherwise obscured.

In this talk, we will discuss recent work on the development of interactive visualization techniques and tools for a variety of needs.

“The AI-Seminar is sponsored by Yahoo!”

May 4th

Speaker: Quoc Le, Stanford University

Host: Ashutosh Saxena

Title : Tera-scale deep learning 

Abstract: Deep learning and unsupervised feature learning offer the potential to transform many domains such as vision, speech, and NLP.  However,
these methods have been fundamentally limited by our computational abilities, and typically applied to small-sized problems.  In this talk, I describe the key ideas that enabled scaling deep learning
algorithms to train a very large model on a cluster of 16,000 CPU cores (2000 machines).  This network has 1.15 billion parameters, which is more than 100x larger than the next largest network reported in the literature.

Such network, when applied at the huge scale, is able to learn abstract concepts in a much more general manner than previously demonstrated. Specifically, we find that by training on 10 million
unlabeled images, the network produces features that are very selective for high-level concepts such as human faces and cats. Using these features, we also obtain significant leaps in recognition
performance on several large-scale computer vision tasks. “The AI-Seminar is sponsored by Yahoo!”

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Scott Sanner, NICTA and the Australian National University

Location: UPSON 315 10:00AM

Host: Thorsten Joachims

Bio: Scott Sanner is a Senior Researcher in the Machine Learning Group at NICTA, having joined in 2007.  Scott earned a PhD from the University of Toronto, an MS degree from Stanford, and a double BS degree from Carnegie Mellon.  Scott's research interests span AI, Machine Learning, and Information Retrieval.  For more information, please visit: http://users.cecs.anu.edu.au/~ssanner/

Title : New Objective Functions for Social Collaborative Filtering

Abstract: This talk examines the problem of social collaborative filtering (CF) to recommend items of interest to users in a social network setting.  Unlike standard CF algorithms using relatively simple user and item features, recommendation in social networks poses the more complex problem of learning user preferences from a rich and complex set of user profile and interaction information.  Many existing social CF methods have extended traditional CF matrix factorization, but have overlooked important aspects germane to the social setting.  We propose a unified framework for social CF matrix factorization by introducing novel objective functions for training.  Our new objective functions have three key features that address main drawbacks of existing approaches: (a) we fully exploit feature-based user similarity, (b) we permit direct learning of user-to-user information diffusion, and (c) we leverage co-preference (dis)agreement between two users to learn restricted areas of common interest.  We evaluate these new social CF objectives, comparing them to each other and to a variety of (social) CF baselines, and analyze user behavior on live user trials in a custom-developed Facebook App involving data collected over five months from over 100 App users and their 37,000+ friends.

This is joint work with Joseph Noel (recommendation algorithm developer), Khoi-Nguyen Tran (Facebook App developer), Peter Christen, Lexing Xie, Edwin V. Bonilla, Ehsan Abbasnejad, Nicolas Della Penna and will appear at WWW 2012.

Paper Link: http://users.cecs.anu.edu.au/~ssanner/Papers/www12.pdf

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Lior Seeman, Cornell University & Joe Halpern, Cornell University

Title : I’m Doing as Well as I Can: Modeling People as Rational Finite Automata

Abstract: We show that by modeling people as bounded finite automata, we can
capture at a qualitative level the behavior observed in experiments.
We consider a decision problem with incomplete information and a dynamically changing
world, which can be viewed as an abstraction of many real-world settings.
We provide a simple strategy for a finite automaton in this setting,
and show that it does quite well, both through theoretical analysis and
simulation. We show that, if the probability of nature changing state
goes to 0 and the number of states in the automaton increases, then this
strategy performs optimally (as well as if it were omniscient and knew
when nature was making its state changes). Thus, although simple, the
strategy is a sensible strategy for a resource-bounded agent to use.
Moreover, at a qualitative level, the strategy does exactly what people
have been observed to do in experiments.

Joint work with Joe Halpern and Rafael Pass.

------------------------------------------------------------------------------------------------------------------

Speaker #2: Joe Halpern, Cornell University

Title: Ambiguous Language and Differences in Beliefs

Abstract: Standard economic models cannot capture the fact that information is often ambiguous, and is interpreted in multiple ways. We propose a framework that can model this, and we show that ambiguity can have important consequences.  One application of this work is to Aumann's classical result showing that players with a common prior cannot agree to disagree:
they cannot agree to a trade, since, roughly speaking, this requires disagreement regarding an expected outcome.  We show that players can agree to disagree in the presence of ambiguity, even if there is a common prior, but that allowing for ambiguity is more restrictive than assuming heterogeneous priors.  This suggests that ambiguity provides a potential explanation for heterogeneous beliefs, while it imposes nontrivial restrictions on the situations that can be modeled, so that it is not the case that ``anything goes'' once we allow for ambiguity.

This is joint work with Willemien Kets.

“The AI-Seminar is sponsored by Yahoo!”

THURSDAY, June 14th

Time & Location: THURSDAY, June 14th: 12:00 - 1:15pm UPSON 5126 - TWO SPEAKERS

Speaker: Yun Jiang, Cornell University

Host: Ashutosh Saxena

Title : Learning Object Arrangements in 3D Scenes using Human Context

Abstract:  We consider the problem of learning object arrangements in a 3D scene. The key idea here is to learn how objects relate to hu- man skeletons based on their affordances, ease of use and reachability. In contrast to modeling object-object relationships, model- ing human-object relationships scales linearly in the number of objects.
We design appro- priate density functions based on 3D spatial features to capture this. We then learn the distribution of human poses in a scene us- ing a variant of the Dirichlet process mixture model, allowing sharing of the density function across same object types.
This allows our algorithm to reason about arrange- ment of the objects in the room. In our exten- sive experiments on 20 different rooms with a total of 47 objects, our algorithm predicted correct placements with an average error of 1.6 meters from ground truth. In arranging five real scenes, it received a score of 4.3/5 compared to 3.7 for the best baseline method.
http://pr.cs.cornell.edu/placingobjects

Speaker: Daniel Ly, Cornell University

Host: Ashutosh Saxena

Title: Co-evolutionary Predictors for Kinematic Pose Inference from RGBD Images

Abstract: Markerless pose inference of arbitrary subjects is a primary problem for a variety of applications, including robot vision and teaching by demonstration. Unsupervised kinematic pose inference is an ideal method for these applications as it provides a robust, training-free approach with minimal reliance on prior information.
However, these methods have been considered intractable for complex models. This paper presents a general framework for inferring poses from a single depth image given an arbitrary kinematic structure without prior training. A co-evolutionary algorithm, consisting of pose and predictor populations, is applied to overcome the traditional limitations in kinematic pose inference. Evaluated on test sets of 256 synthetic and 52 real images, our algorithm shows consistent pose inference for 34 and 78 degree of freedom models with point clouds containing over 40,000 points, even in cases of significant self-occlusion. Compared to various baselines, the co-evolutionary algorithm provides at least a 3.5-fold increase in pose accuracy and a two-fold reduction in computational effort for articulated models.

“The AI-Seminar is sponsored by Yahoo!”