Pictorial structure (PS) models are extensively used for part-based recognition of scenes, people, animals and multi-part objects. To achieve tractability, the structure and parameterization of the model is often restricted, for example, by assuming tree dependency structure and unimodal, data-independent pairwise interactions. These expressivity restrictions fail to capture important patterns in the data. On the other hand, local methods such as nearest-neighbor classification and kernel density estimation provide non-parametric flexibility but require large amounts of data to generalize well. We propose a simple semi-parametric approach that combines the tractability of pictorial structure inference with the flexibility of non-parametric methods by expressing a subset of model parameters as kernel regression estimates from a learned sparse set of exemplars. This yields query-specific, image-dependent pose priors. We develop an effective shape-based kernel for upper-body pose similarity and propose a leave-one-out loss function for learning a sparse subset of exemplars for kernel regression. We apply our techniques to two challenging datasets of human figure parsing and advance the state-of-the-art (from 80% to 86% on the Buffy dataset), while using only 15% of the training data as exemplars.

In many image and video collections, we have access only to partially labeled data. For example, personal photo collections often contain several faces per image and a caption that only specifies who is in the picture, but not which name matches which face. Similarly, movie screenplays can tell us who is in the scene, but not when and where they are on the screen. We formulate the learning problem in this setting as partially-supervised multiclass classification where each instance is labeled ambiguously with more than one label. We show theoretically that effective learning is possible under reasonable assumptions even when all the data is weakly labeled. Motivated by the analysis, we propose a general convex learning formulation based on minimization of a surrogate loss appropriate for the ambiguous label setting. We apply our framework to identifying faces culled from Web news sources and to naming characters in TV series and movies. We experiment on a very large dataset consisting of 100 hours of video, and in particular achieve 6% error for character naming on 16 episodes of LOST.

wikiSearch, is a search engine customized for the Wikipedia corpus but with design features that may be generalized to other search systems. Its features enhance basic functionality and enable more fluid interactivity while supporting both workflow in the search process and the experimental process used in lab testing.

Abstract Reading is a common everyday activity for most of us. In this article, we examine the potential for using Wikipedia to fill in the gaps in one's own knowledge that may be encountered while reading. If gaps are encountered frequently while reading, then this may detract from the reader's final understanding of the given document. Our goal is to increase access to explanatory text for readers by retrieving a single Wikipedia article that is related to a text passage that has been highlighted. This approach differs from traditional search methods where the users formulate search queries and review lists of possibly relevant results. This explicit search activity can be disruptive to reading. Our approach is to minimize the user interaction involved in finding related information by removing explicit query formulation and providing a single relevant result. To evaluate the feasibility of this approach, we first examined the effectiveness of three contextual algorithms for retrieval. To evaluate the effectiveness for readers, we then developed a functional prototype that uses the text of the abstract being read as context and retrieves a single relevant Wikipedia article in response to a passage the user has highlighted. We conducted a small user study where participants were allowed to use the prototype while reading abstracts. The results from this initial study indicate that users found the prototype easy to use and that using the prototype significantly improved their stated understanding and confidence in that understanding of the academic abstracts they read.

The data deluge is beginning to have an effect on libraries and archives. As custodians of the scholarly record, libraries and archives are being asked to play an active role in long-term digital preservation in both science and the humanities. A report to the National Science Foundation from the fall 2006 ARL Workshop on the role of academic libraries in the digital data universe states that "the group found that research and academic libraries need to expand their portfolios to include activities related to storage, preservation and curation of digital scientific and engineering data." (To Stand, p 42) One of the major trends in this area is the notion of partnerships, of considering the full set of skills necessary to preserve data for the long term and recognizing that a single group or discipline does not have expertise in all aspects of digital preservation. Libraries and archives provide expertise in information management, organization and accessibility. Computer scientists and engineers provide expertise in the portfolio of technologies required to support digital preservation. Domain scientists and humanities scholars provide expertise in the content of the data to be preserved. In order to be effective, these groups must work together. This poster describes three such collaborations based at the Texas Advanced Computer Center, the San Diego Supercomputer Center, and Indiana University.

Over the past 7 years, the San Diego Supercomputer Center has worked to produce the highest possible performance file systems available to the National Science Foundation community in the USA. Most of this was done with GPFS, IBM's parallel file system, but several distinctly different configurations were designed and implemented with numerous lessons learned in the process. All of these systems provided transfer rates in the multiple GB/s range. In this paper, we detail the configurations and their intended modes of operation and, as much as possible, show the resulting performance. We attempt to describe the advantages and disadvantages of each approach with an emphasis on the implications for future systems

Extracting morphemes from words is a nontrivial task. Rule based stemming approaches such as Porter's algorithm have encountered some success, however they are restricted by their ability to identify a limited number of affixes and are language dependent. When dealing with languages with many affixes, rule based approaches generally require many more rules to deal with all the possible word forms. Deriving these rules requires a larger effort on the part of linguists and in some instances can be simply impractical. We propose an unsupervised ngram based approach, named Swordfish. Using ngram probabilities in the corpus, possible morphemes are identified. We look at two possible methods for identifying candidate morphemes, one using joint probabilities between two ngrams, and the second based on log odds between prefix probabilities. Initial results indicate the joint probability approach to be better for English while the prefix ratio approach is better for Finnish and Turkish.

We investigate here, both theoretically and by demonstration, scaling file storage to the very widest possible extents. We use IBM's GPFS file system, with extensions developed by the San Diego Supercomputer Center in collaboration with IBM. Geographically, the file system extends across the United States, including Pittsburgh, Illinois, and San Diego, California, with the TeraGrid 40 Gb/s backbone providing the wide area network connectivity. We show the results from two demonstrations, at each of the past two supercomputing conferences, SC03 in Phoenix, Arizona, and SC04 in Pittsburgh, Pennsylvania. The second demonstration was purposely designed to presage an intended production facility across the National Science Foundation's TeraGrid.

In this paper we describe the evolution of Global File Systems from the concept of a few years ago, to a first demonstration using hardware Fibre Channel frame encoding into IP packets, to a native GFS, to a full prototype demonstration, and finally to a production implementation. The surprisingly excellent performance of the Global File Systems over standard TCP/IP Wide Area Networks has made them a viable candidate for the support of Grid Supercomputing. The implementation designs and performance results are documented within this paper. We also motivate and describe the authentication extensions we made to the IBM GPFS file system, in collaboration with IBM. In several ways Global File Systems are superior to the original approach of wholesale file movement between grid sites and we speculate as to future modes of operation.