Joydeep Ghosh - US grants

An integrated approach towards the synthesis of high-bandwidth, low- latency interconnection networks is pursued that focuses on the interdependence of system architecture, VLSI characteristics and packaging technology. New designs of one-sided crosspoint switching modules that can provide several paths between communicating ports in multichip switches are being studied. An architecture solution is posed for the simultaneous switching noise problem and that has been traditionally tackled at the circuit design and packing levels. Novel schemes for fast arbitration and scheduling through the use of parallel bus-controllers are analyzed under a variety of message traffic conditions. An alternative approach uses artificial neural networks for contention resolution. Trade-offs between switch cost, performance and reliability are studies under these workloads for various chip sizes. Topological testing techniques are used for drastic reduction of test time. The research findings are being incorporated into a set of software tools recently developed at UT Austin for the computer-aided design of networks. Recent technological advances have prompted a resurgence of interest in cross bar-based designs in the industry, despite the requirement of O (N2) switches. Indeed, by considering current VLSI and packaging technology, it can be shown that power dissipation, current driving abilities and pinout limitations easily outweigh circuit density demands of switching networks. This research aims to lay the foundations for the cost-effective design of 128-1024 port crossbars with a sustained bandwidth of 1 GBytes per channel, that can form the core of the next generation of massively parallel machines and switching networks.

9307632 Sandberg Nonlinear transformation and recognition of continuous-time signals or signal sequences is fundamental to a wide range of cognitive processes. This work aims to build a comprehensive understanding of the processing of spatio-temporal signals. It is founded on recent results obtained by the proposer showing that very large classes of continuous functionals and shift-invariant functional maps can be uniformly approximated by certain conceptually simple neural-like structures. These structures are the Function Space Neural Networks (FSNNs) that involve a preprocessing linear operation stage such as convolution with suitable kernel functions, followed by a network of nonlinear cells. The project shall address key issues pertaining to the design and use of FSNNs. These include determination of suitable kernel functions, network size, connectivity patterns and form of nonlinearity for different problems classes, effectiveness of alternate learning algorithms and their convergence rates, and techniques for constructive/destructive network growth. Anew call of FSNNs based on higher order networks that have proved very effective for static classification, will also propose be evaluated. Theoretical studies shall be supplemented by extensive simulations using several suites of spatio-temporal signals ranging form low-dimensional artificial patterns to a set of over 1000 short duration signals representing actual passive sonar returns form underwater biologics. This work shall try to raise the understanding of the capabilities and usability of artificial neural network structures for the processing of spatio-temporal signals to a level comparable to that achieved at present for (static) multilayered feedforward networks. ***

9900353
Ghosh

Simultaneous and coordinated use of multiple learning components is often needed for satisfactory and robust solutions to difficult problems involving learning, adaptation, recognition/classification and dynamic control. In popular multi-learner approaches such as ensembles, Bayesian model averagers and mixtures of experts, each component model tries to solve (possibly localized versions of) the same task. However, in practice, one is often faced by a series of (possibly related) tasks, or a task whose nature changes substantially with time due to nonstationarities, new data/sensors or new domain knowledge. Moreover, these tasks are solved at different points in time, and at any given time, we may know about what future tasks will need to be solved.

This project will study how existing "support" models can be leveraged to help solve new, and possible related tasks. It will focus on classification problems, though many of the approaches extend to a much wider variety of tasks. Issues to be examined include: (i) how to select from among several support models and measure their relevance vis a vis the new task; (ii) transfer/reuse mechanisms that only use input/output behavior of support models, and (iii) mechanisms that exploit internal structural information of support models. Certain special issues that arise when dealing with sequence classifiers will also be investigated. The main goal is to exploit knowledge encapsulated in support models for quicker, more accurate learning with fewer new training examples, along with better understanding of the new problem.

This project will also establish a benchmarking facility in the proposed areas for the benefit of the research community, based on engineering related datasets. This work will facilitate the application of more versatile and powerful multi-component approaches to engineering smart, adaptive systems. It will also create an evolving knowledge base of support models that are developed collaboratively and are reusable for future tasks.
***

ABSTRACT

"ACTIVE STEREOSCOPIC VISUAL SEARCH DRIVEN BY NATURAL SCENE STATISTICS"

Alan C. Bovik, Lawrence K. Cormack, J. Ghosh

The primary thrust of this proposal is to develop methods based on the natural statistics of stereoscopic images that will enable the design and implementation of the next generation of foveated, fixating machine vision systems that are capable of efficient and intelligent visual search, by exploiting and applying knowledge about human fixation and search mechanisms.
We summarize the intention of our proposal via the following key goals:
Goal 1: To develop a quantitative description of human active stereo vision as a function of natural scene statistics in a variety of three-dimensional visual search and learning tasks. Our emphasis will be on developing statistical models of stereo primitives that attract low-level visual attention based on a unique and in-depth statistical analysis. We feel that statistical models based on natural scene statistics have a very good chance of succeeding where deterministic models have failed.
Goal 2: To train a state-of-the-art foveated, fixating active computer vision system (named FOVEA) to search and to learn to search 4-D (space-time) scenes. To do this, back-end artificial neural networks trained on telepresent human search patterns will be used. The statistical models and extracted statistical stereoprimitives discovered as part of the research in Goal 1 will be used as a priori knowledge to improve the configuration and learning of the networks. We envision that these experiments will result in smart active machine vision protocols for exploring, searching, and interacting with 4-D environments, while giving new insights into visual cognitive processes.

This award will provide for the development of a comprehensive system that can efficiently and intelligently extract, analyze and manage very large hyperspectral datasets used for classifying a large variety of land covers in environmentally sensitive ecosystems.

Hyperspectral data provide unprecendented spectral resolution which can translate to far superior characterization of remotely sensed areas, but pose significant challenges because of the large data volumes, high dimensionality, little labelled data and large number of potential land cover types or classes. These challenges are being addressed by new adaptive feature space reduction methods that exploit spectral correlations, by semi-supervised and active learning methods for dealing with small training sets, and by knowledge reuse and transfer mechanisms that adapt models developed for one area to new regions with related characteristics. In parallel, a knowledge repository that helps rapidly identify the most pertinent features/classes for a given area, will be built to substantially reduce data storage requirements and processing time.

This inter-disciplinary project requires tight interaction between data acquisition and processing/analysis, and will provide insights for other engineering problems as well. The visual nature of results from analysis of remotely sensed data make it a powerful modality of introducing the general population to issues of broad concern, such as the impact of global warming and disaster management. Finally, the knowledge transfer mechanisms will be useful for rapidly adapting existing solutions to somewhat different but related problems, thus substantially increasing the utility of existing point solutions in several application domains.

This research addresses three key issues pertaining to the clustering of large, complex datasets. First, a unifying view of model-based clustering will be developed to form a theoretical basis for understanding and comparing a wide range of existing clustering algorithms for complex data. This view will be systematically explored to develop improved algorithms for specific applications. Second, complexity arising from domain constraints on balancing and dealing with incrementally acquired non-stationary data, such as newsfeeds, is addressed via adaptive clustering techniques. Finally, methods for obtaining a single consensus solution given multiple clustering results are investigated. Such methods will facilitate distributed data mining under severe restrictions on data sharing due to privacy and other constraints. Benchmarks for the proposed research areas will be developed and made available to the research community. Further information about the project is available on the project web site http://www.lans.ece.utexas.edu/scalclust.html. Broader impacts of this project also include outreach activities to high school students and freshmen students. Demonstration modules that illustrate data analysis issues will be designed for students and enable them to use case studies resulting from this work and gain understanding of clustering techniques.

Cluster analysis is an indispensable tool in the data miner's arsenal which enables one to understand the structure of the data while conducting exploratory data analysis. Recent times have seen increased occurrences of bi-modal and multi-modal data that manifest themselves as two-dimensional matrices and higher-dimensional tensors. Co-clustering is becoming an increasingly popular technique for exploratory analysis of such data, and has been successfully applied in wide range of areas, including web mining, natural language processing, image and video content analysis, recommender systems, and bioinformatics . The broad goal of this project is to develop sound, theoretical formulations of varied types of co-cluster analyses so that co-cluster analysis becomes an indispensable and efficient tool in the exploratory analysis of bi-modal and multi-modal data.

This research focuses on extending co-cluster analysis to include multi-dimensional tensors where one desires to cluster on more than two modes simultaneously, and matrix data with added row and column attributes such as those describing networked knowledge structures or multiple interlinked tables. This will enable co-clustering to reach a much wider class of applications and also make it computationally practical.

In order to broaden the impact of this project, the principal investigators are jointly organizing workshops that foster and promote research on various aspects of co-cluster analysis. Data, papers and software developed under this project will be shared with the scientific community via the project Web site (http://hercules.ece.utexas.edu/~ghosh/scalclust.html). Finally, as part of community outreach, the investigators plan to design outreach modules that illustrate data analysis concepts and capabilities at levels appropriate for high school students as well as for freshmen students.

Intellectual Merits.
The characterization of land cover and usage over large geographical regions, as well as the near/long-term forecasting of changes in land use, is a key problem in geo-informatics that is particularly important for regions that are subject to rapid ecological changes or urbanization. At present, the data and knowledge required for detailed and accurate characterization is scattered across both traditional (GIS) spatial data sources and across remotely sensed data, and their associated models, none of which inter-operate well. This research will develop a comprehensive framework for efficient and accurate mapping, monitoring and modeling of land cover and changes in usage over large regions. This endeavor involves three complementary activities: (i) large scale classification of remote sensing imagery using advanced learning methods, including transfer learning, active learning and manifold based data descriptors; (ii) next-generation spatial modeling using ensembles for forecasting land transformations; and (iii) integration of GIS and remote sensing data by distributed, privacy aware learning, integrating taxonomies obtained from different data sources and portal building. A plan of interaction with various stakeholders is proposed to ensure that the results are meaningful and actionable. This project will result in substantial advances in analysis of remotely sensed data over extended regions and lead to a substantial reduction in the uncertainty of long-term forecasts of change. Concurrently, the chosen application domain will also provide a concrete setting that motivates several new data mining problems, leading to new algorithmic formulations and solutions that benefit the broader data mining community.

Broader Impacts.
This project is designed to have many, diverse broader impacts. First is the involvement of application scientists in the remote sensing and modeling communities who will benefit from advanced methods in machine learning. The research results will be brought into the classroom through new graduate courses. Popular science lectures for middle and high school are also planned since the subject matter and results can be conveyed meaningfully to this audience in a visual way that emphasizes issues of broader concern, such as the impact of ecological changes and urban sprawl. Two project-wide workshops are proposed that will also involve stakeholders (e.g., planners) who would directly benefit from the results and provide valuable feedback. A portal will be created in year 3 to provide access to data, code and toolkits produced by the project. Results will be disseminated in each of the three main disciplines represented within the project through scholarly publications. Finally, tools will be developed so that they may eventually be incorporated into Commercial Off The Shelf software, such as GIS and remote sensing software.

Several modern data mining applications involve predictive modeling on large
amounts of multi-relational data with added structures such as product
hierarchies or social networks among customers. The broad goal of this proposal
is to develop a comprehensive framework for predictive modeling on large,
heterogeneous, multi-relational data based on "Simultaneous Decomposition and
Prediction" (SDaP) approaches that iteratively partition the problem into more
homogeneous and manageable pieces while concurrently building multiple
predictive models, one for each piece. Such approaches lead to simpler and more
accurate solutions. The proposed algorithmic strategies that determine how many
models to learn and where they should apply, which data to discard and which to
keep, how to learn multiple related tasks defined on multi-modal data, and how
to scalably implement the solutions on distributed computers, provide practical
solutions to certain real-world problems for which current learning and data
mining techniques are severely lacking. Application domains of ecology, bio-
informatics, market research and web mining are specifically identified and
targeted.

There are two broad research impacts of the proposed project: (a) it further
vitalizes the research in data mining towards better algorithms for predictive
modeling on rich and heterogeneous multi-modal data, and (b) provides and
promotes the SDaP approach as a fundamental data analysis tool across multiple
disciplines. The PI will organize a workshop and offer a tutorial at major data
mining conferences to foster and promote research on various aspects of SDaP
analysis. Moreover, the curated complex datasets and software developed under
this project will be shared with the scientific community via a public web site
as part of the proposed one-of-a-kind multi-relational data benchmarking
facility. The PI will further develop a novel graduate course on Modeling and
Analysis of Complex Data. Outreach modules that illustrate data analysis
concepts and capabilities at levels appropriate for pre-college students will
also be developed. For further information see the project web site at the URL:
http://www.ideal.ece.utexas.edu/projects/sdap/

Determining the perceptual quality of video transmitted through complex networks and viewed on heterogeneous platforms, from cell phones to Internet-based television, is a key problem for the YouTube generation. It is also central to a variety of vision applications including face detection, face recognition and surveillance. Video is subject to numerous distortions: blur, noise, compression, packet/frame drops, etc. Quality assessment is non-trivial when an undistorted video is not available, and unsolved for multiple distortion types and in distributed, non-stationary viewing environments.

This project designs and creates intelligent video "quality agents" that learn how to determine perceptual video quality in heterogeneous networks, and assesses its impact on decision tasks such as face detection and recognition, all without the benefit of reference videos. It uses statistical properties of natural scenes, perceptual principles, machine learning, and intelligent adaptive agent collectives to handle videos simultaneously impaired by multiple distortion types. A primary application is novel face-salient quality assessment agents and quality-aware face detection algorithms. Multiple, co-operative video and face quality agents are trained using active learning based feedback mechanisms on mobile devices. This project yields adaptive, robust video Quality of Service assessment in real-life networks and provides new insights into human visual quality perception and visual distortion detection. The research team also creates two large, unique video quality databases: (a) A Mobile Video Quality Database of raw and distorted mobile videos and (b) A Distorted Face Database of undistorted and distorted face images, as gold standards for research and development in this area.

Determining preferences, or identifying and ordering items of most interest or relevance based on very limited information, is a fundamental problem in many disciplines. While perhaps most obvious in search, the problem shows up in areas as diverse as economics and health informatics. This project is developing a broad methodology to address challenging learning problems that involve ordering, including determining top choices for recommendation, multi-label classification, and learning to rank-order a set of query results. The key insight is that if the objects to be ordered are given numeric scores, only the relative values of these scores affect their ranking, and not the actual values. This project is using this insight to develop new and better learning algorithms for ranking.

Specifically, the project is developing methods that can efficiently optimize over all possible monotonic (that is, order preserving) transformations of scores. Since these scores become the target values for regression, this class of approaches is being called monotonic retargeting. A systematic way of alternating between readjusting scores and updating the regression model is being developed, with nice properties for scalability and distributed implementation, as well as strong convergence guarantees. Themes common to different types of preference learning or ranking studies are being identified to help bring together the diverse communities, including students, that work on this topic. This wide coverage is possible as it easy to relate to the need to determine priorities and make choices in various walks of life. The applications and impacts of the project are expected to be wide and diverse as well.

As the adoption of electronic health records (EHRs) has grown, EHRs are now composed of a diverse array of data, including structured information (e.g., diagnoses, medications, and lab results), molecular sequences, unstructured clinical progress notes, and social network information. There is mounting evidence that EHRs are a rich resource for clinical research, but they are notoriously difficult to leverage because of their orientation to healthcare business operations, heterogeneity across commercial systems, and high levels of missing or erroneous entries. Moreover, the interactions among different data sources within an EHR are challenging to model, hampering our ability to leverage traditional analytic frameworks. In recognition of this problem, various efforts have been undertaken to transform EHR data into concise and meaningful concepts, or phenotypes. Yet, to date, these efforts have been ad hoc and labor intensive, resulting in specific phenotypes for specific environments; e.g., type 2 diabetes in the EHR system at Vanderbilt University Medical Center (VUMC). There is an urgent need for scalable phenotyping methods, but several major challenges must be addressed, including: a) patient representation, b) high-throughput phenotype generation from EHRs, c) expert-guided phenotype refinement, and d) phenotype adaptation across institutions. The goal of this project is to address these challenges by developing a general computational framework for transforming EHR data into meaningful phenotypes with only modest levels of expert guidance. The PIs will develop novel courses on Healthcare Analytics as a Massive Open Online Course (MOOC) that covers cross-disciplinary topics at the confluence of computer science and medical informatics, while embellishing existing graduate courses on biomedical informatics. The PIs plan to deliver tutorials and organize workshops at relevant computer science and medical informatics conferences with the goal of sharing research results and developing a community. The PIs will develop outreach modules that focus on freshmen and under-represented students, as well as educational sessions for clinical researchers who are currently performing phenotyping in academic medical centers. Thus, the project has a significant component the integrates research and education as well as providing for new scientific insights.

In support of this goal, the team plans to represent and analyze EHR data as inter-connected high-order relations i.e. tensors (e.g. tuples of patient-medication-diagnosis, patient-lab, and patient-symptoms). The proposed analytic framework generalizes several existing data mining methodologies, including dimensionality reduction, topic modeling and co-clustering, which all arise as limited special cases of analyzing second order tensors. It will also enable flexible refinement of candidates to adapt phenotypes from one healthcare institution to another, and will incorporate feedback from domain experts. The accompanying suite of algorithms and methods will enable the automation of high-throughput phenotype generation, refinement, adaptation and applications, in a broad range of health informatics settings and across multiple institutions. This project will integrate biomedical informaticists, computer scientists, and clinical experts. The significance of the resulting phenotypes in diverse clinical applications, including: a) cohort construction, where case and control patients are identified with respect to specific phenotype combinations; b) genome wide association studies (GWAS), where target phenotypes of patients are tested against DNA sequence variation for significant statistical associations; and c) clinical predictive modeling, where a model is developed to predict target phenotypes or diseases will be demonstrated. The framework will be developed with public accessible data from MIMIC-II and CMS and validate in real clinical environments at Northwestern Memorial Hospital and VUMC through several high-impact disease targets (including hypertension, type 2 diabetes, hypothyroidism, atrial fibrillation, rheumatoid arthritis, and multiple sclerosis). Additionally, the methodologies developed through this project will be integrated into existing software platforms that support the representation of EHR-derived phenotypes, but lack a data-driven component for the generation and refinement of candidates. Overall, the proposed framework is expected to have a major impact on translational clinical research including clinical trial design, predictive modeling, epidemiology studies and clinical decision support.