Abstracts

• Christos Faloutsos, Computer Science, Carnegie Mellon Univesity

  Title: Graph Mining: Laws, Generators and Tools

  How do graphs look like? How do they evolve over time? How can we generate realistic-looking graphs? We review some static and temporal 'laws', and we describe the ``Kronecker'' graph generator, which naturally matches all of the known properties of real graphs. Moreover, we present tools for discovering anomalies and patterns in two types of graphs, static and time-evolving. For the former, we present the 'CenterPiece' subgraphs (CePS), which expects $q$ query nodes (eg., suspicious people) and finds the node that is best connected to all $q$ of them (eg., the master mind of a criminal group). We also show how to compute CenterPiece subgraphs efficiently. For the time evolving graphs, we present tensor-based methods, and apply them on real data, like the DBLP author-paper dataset, where they are able to find natural research communities, and track their evolution.

  Finally, we also briefly mention some results on influence and virus propagation on real graphs, as well as on the emerging map/reduce approach and its impact on large graph mining.

• Usama M. Fayyad, Chief Data Officer and Executive Vice President, Yahoo Inc.

  Title: Towards Inventing the New Sciences of the Internet -- Searching, Marketing, and the Power of Social Networks

  As the Internet continues to change the way we live, find information, communicate, and do business, it has also been taking on a dramatically increasing role in marketing and advertising. The uniqueness of the internet as a measureable interactive medium makes dealing with large volumes of data and its analysis and modeling an essential mission-critical part of running the on-line business. Furthermore, the most basic fundamental concepts underpinning many of the applications that are becoming popular are based on a shallow understanding of what they are and how they can be understood. The larger issues surrounding the Internet as a technology that is ubiquitous in our lives, yet little is understood at the scientific level, is creating a fundamental set of fundamental challenges in defining and understanding many of the basics of the Internet: Community Systems, Search and Information Navigation, Computational Advertising, and the new Microeconomics of the Web. As an example, in scarcely a decade, web search has gone from simply scaling traditional information retrieval, to a plethora of new opportunities that are changing marketing as we know it. In this talk we begin by reviewing this progress, pointing out that web search is no longer a purely computer science problem. We then hint at the role of other disciplines in this ongoing revolution and the challenges ahead of us. I will use some challenges in search and how it is used, as well some recent developments at Yahoo! in terms of how we think Search will evolve. This will lead us into a discussion of the new Yahoo! Research organization and its aims: inventing the new sciences underlying what we do on the Internet, focusing on areas that have received little attention in the traditional academic circles and industrial research labs. Some illustrative examples will be reviewed to make the ultimate goals more concrete. Particular focus will be provided around data-driven capabilities and how data mining plays a unique role in driving many initiatives.

• Michael Kearns, Computer and Information Science, University of Pennsylvania

  Title: Collective Behavior and Machine Learning

  I will begin by describing an ongoing and extensive series of human subject experiments, conducted at Penn, in collective decision-making and problem-solving over networks from local information. These controlled experiments have shed light on the relationships between network structure, the problem being solved, locality of information, and incentives.
  The goal of using the experimental data to draw generalizations and predictions about future experiments also points out some interesting algorithmic and modeling challenges for machine learning methods. In the second part of the talk I will describe a recent theoretical framework we have developed for learning from collective behavior.
  This talk describes joint work with Stephen Judd, Jennifer Wortman, Jinsong Tan, Siddharth Suri, and Nick Montfort.

• Tomaso A. Poggio, Computer Science and Artificial Intelligence Laboratory, MIT

  Title: Visual Recognition in primate cortex: from Neuroscience to a new AI?

  Understanding the processing of information in our cortex is a significant part of understanding how the brain works, arguably one of the greatest problems in science today. In particular, our visual abilities which seem effortless to us are computationally amazing: computer science is still far from being able to create a vision engine that imitates them. Thus, visual cortex – and the problem of the computations it performs -- may well be a good proxy for the rest of the cortex and indeed for intelligence itself.

  In the talk I will review our recent work on developing a hierarchical feedforward architecture for object recognition based on the anatomy and the physiology of the primate visual cortex. The main goal of the model is to predict properties of neurons in a series of visual areas from primary visual cortex to inferotemporal cortex. The model helps to interpret the results of experiments and to plan new ones. I will briefly describe some of the related collaborative work with cortical physiologists. Somewhat surprisingly, the model performs at the level of human subjects for short presentations on a difficult natural image recognition task. Furthermore, the performance on several databases of complex images was at the level of the best available computer vision systems. I will sketch current work aimed to a) develop a mathematical theory of such hierarchical architectures for learning b) extend the model to incorporate attentional control and c) extend the model to deal wit the recognition of objects and actions in videos. We are also beginning to work towards an architecture that deals with learning from very few examples and could eventually solve a broad spectrum of image inference tasks beyond image categorization – thus going beyond perception and towards cognition.

  The emerging thesis of this talk is that neurally plausible computational models are beginning to provide powerful new insights into the key problem of how the brain works, and of how to implement learning and intelligence in machines. Thus the neurosciences may begin to exert a strong pull on the evolution of artificial intelligence.