To Preserve Or NotTo Preserve Or Not To Preserve? The Challenges inThe Challenges in Appraising Electronic Recordsect o c eco ds
Peter Bajcsy, PhD- Research Scientist, NCSA- Adjunct Assistant Professor ECE & CS atAdjunct Assistant Professor ECE & CS at UIUC- Associate Director Center for Humanities, Social Sciences and Arts (CHASS), Illinois Informatics Institute (I3), UIUC
National Center for Supercomputing ApplicationsUniversity of Illinois at Urbana-Champaign
Date: January 21st, 2009
Acknowledgement
• This research was partially supported by a National Archive and Records Administration (NARA) supplement ( ) ppto NSF PACI cooperative agreement CA #SCI-9619019 and NCSA Industrial Partners.The ie s and concl sions contained in this doc ment• The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the National Archive and Records Administration, or the U.S. government.
• Contributions by: Peter Bajcsy Kenton McHenry Rob• Contributions by: Peter Bajcsy, Kenton McHenry, Rob Kooper, Michal Ondrejcek, William McFadden, Sang-Chul Lee, David Clutter and Alex Yahja
Imaginations unbound
Outline
• IntroductionStakeholders• Stakeholders
• Conceptual Challenges• Some Open Problems• Research Examples Illustrating OpenResearch Examples Illustrating Open
Problems• Summary Observations and Future• Summary, Observations and Future
Vision
Introduction• Two Trends in the Context of Decision Processes
(Government, Medical, Natural Disasters, …) • Decision processes are moving from paper based
to electronic record based (~ computer assisted decision processes)decision processes)
• Electronic records depend on rapidly changing information technologyinformation technology
• Decisions are optimal depending on knowledge• Any learning from electronic records depends on• Any learning from electronic records depends on
preservation and reconstruction of the records, as well as on quality and granularity of the information
National Center for Supercomputing Applications
well as on quality and granularity of the information
Fundamental Problems
• Limited learning from historical records todaytoday• It is often due to missing information and
high uncertainty/ low quality of historicalhigh uncertainty/ low quality of historical records.
• Lack of understanding how to preserve and reconstruct data and decision processes.• It is due to insufficient
forecasting/simulation capabilities.
National Center for Supercomputing Applications
To Be Preserved!Digital representation of i f ti
Preservationinformation & knowledge
Information transfer ?
Imaginations unbound
AGENCY ARCHIVES
Motivation• The problems related to preservation of electronic records
are only going to become more serious• Information becomes more heterogeneous and complexInformation becomes more heterogeneous and complex
• More data types• Higher dimensional data
N fil f t• New file formats• Volumes of electronic records have been increasing and will
continue to grow• The model of a paperless office (4 years of Bush’s email > 8
years of Clinton’s email)• The paradigm shift to eScience
• Digital information technology has been changing faster than any previous preservation media
• The time scale of electronic media is ephemeral in comparison p pwith paper or clay tablets
Imaginations unbound
Example of Preservation Needs in Medicine
• Short term:• Medical practice requires comparing patients’
records acquired today with the patients’ d f 5 10 50 70 i d trecords from 5, 10, 50, or 70 years in order to
assess functional, structural or low level biological changes due to diseasesbiological changes due to diseases, treatments and/or aging.
• Long term:Long term:• Genealogy studies compare data sets over
several hundreds and thousands of years
National Center for Supercomputing Applications
y
Who Are the Stakeholders?• Multiple institutions and organizations are active in the area• Multiple institutions and organizations are active in the area
of medical record preservation• National Library of Medicine (NLM) y ( )• Research Information Network (RIN) • Medical Research Council (MRC) in UK • National Archives and Record Administration (NARA)
• Identified common goals:S l i t t d t di ll ti• Seamless, uninterrupted access to expanding collections of biomedical data, medical knowledge, and health information
• Preserve medical record collections in highly usable forms and contribute to comprehensive strategies for preservation of biomedical information in the U S and
National Center for Supercomputing Applications
preservation of biomedical information in the U.S. and worldwide.
Other Stakeholders• Government agencies
• Prediction of patterns signaling natural disasters b d hi t i l tbased on historical measurements
• Detection of terrorist attacks based on past experienceexperience
• Learning about other planets from past space shuttle missions
• Preservation of cultural heritage• Companies
P ti f i i d i d• Preservation of engineering drawings and architectural designs – Boeing, John Deere, GM
• Preservation of simulation results – Caterpillar, Fordp ,• Backward compatibility of hardware/software - GE
Imaginations unbound
NARA as One of the Key Stakeholders• According to The Strategic Plan of The
National Archives and Records Ad i i t ti 2006 2016 “P i thAdministration 2006–2016. “Preserving the Past to Protect the Future”• “Strategic Goal: We will preserve and• Strategic Goal: We will preserve and
process records to ensure access by the public as soon as legally possible” p g y p• “D. We will improve the efficiency with
which we manage our holdings from th ti th h d l d th hthe time they are scheduled through accessioning, processing, storage, preservation, and public use.”preservation, and public use.
Conceptual Challenges• Learning Requires Reusing Electronic Records
• How to enable and support preservation and reconstruction of electronic records?reconstruction of electronic records?
• Advancing Sensors and Instruments Leads to New Types of High Dimensional Data and Large VolumesTypes of High Dimensional Data and Large Volumes• How to design preservation methodologies that
scale well? • Process to Enable Learning over Time from
Electronic Records Requires Large Financial InvestmentsInvestments• How to minimize computational hardware,
software and storage cost and maximize the
National Center for Supercomputing Applications
software, and storage cost and maximize the amount of preserved information?
What Are The Key Open Problems?
Imaginations unbound
Some Open Problems -> Intellectual Merit• Appraisal Methodology
• Appraisal by Visual Exploration• Support of Appraisals by Enabling ComparisonsSupport of Appraisals by Enabling Comparisons• Scalability of Appraisals with Increasing Heterogeneity of
Information, Dimensionality of Data and Volume of Electronic RecordsRecords
• Support of Archival Decisions• Simulate Preservation Costs as a Function of Information
G l it d I f ti T h lGranularity and Information Technology • Optimal Utilization of Computational and Human Resources
• Automation of Processing for Preservation g• Discovery of Relationships Among Electronic Records• Information Preserving Conversions of Electronic Records• Sampling Authenticity and Integrity Verification of a Collection of• Sampling, Authenticity and Integrity Verification of a Collection of
Temporally Changing RecordsImaginations unbound
Broader ImpactsProcess to Enable Learning Over Time
+$ KnowledgeElectronic Records
-$
Optimal Decision Making
National Center for Supercomputing Applications
Concrete Research Examples Illustrating Open Problemsp
Imaginations unbound
Open Problems Related to AppraisalOpen Problems Related to Appraisal Methodology
1. Appraisal by Visual Exploration2. Support of Appraisals by Enabling Comparisons3. Scalability of Appraisals with Increasing Heterogeneity of
Information, Dimensionality of Data and Volume of Electronic Records
Imaginations unbound
Records
Definition of Appraisal in Archival Context
• Appraisal -- the process of determining the value and thus the final disposition of Federal records making them eitherthe final disposition of Federal records, making them either temporary or permanent. • See http://www.archives.gov/records-p g
mgmt/initiatives/appraisal.html• The basis of appraisal decisions may include
th d ' d t t• the records' provenance and content, • the records' authenticity and reliability, • the records‘ order and completeness• the records order and completeness, • the records‘ condition and costs to preserve them, and • the records‘ intrinsic valuethe records intrinsic value
Imaginations unbound
Open Problem 1: Appraisal by Visual ExplorationExploration
• How to visualize the transition from raw data to information?• Raw data (Byte stream) -> Information
• How to encode and represent heterogeneous information for visual exploration and for computer assisted operations?
0F0 ->(R.G,B)->GREEN
visual exploration and for computer-assisted operations?• Encoding (e.g., shape consisting of a set of Bezier
curves is encoded by a set of straight lines)• Representation (e.g., colors are represented by an
ordered sequence of intensity values from all bands)H t i t ti f i l l ti ?• How to summarize representations for visual exploration?• Frequency of occurrence of primitives• Local and global summarizations• Local and global summarizations
Imaginations unbound
Example: Adobe Portable Document Format (PDF)Format (PDF)
• Why PDF? - PDF is just an example of a container• Office environment (Adobe PDF PS MS Word HTML )Office environment (Adobe PDF, PS, MS Word, HTML, …)• Satellite measurements (HDF, netCDF, …)
3D
Adobe Library 6.0
Movie
Ad b Lib 7 0Adobe Library 7.0
Imaginations unbound
Exploration of PDF Documents Using PDF ViewerViewer• PDF Viewer presents information as a set of pages with
their layoutstheir layouts• PDF Viewer renders layers of internal objects
(components) and hence only the top layer is visible
Needed Exploration of PDF Componentsp p• There is no support for archival appraisals that would
include visual exploration of components in a document (a container of components)
Needed viewers for appraisal analyses that present• Needed viewers for appraisal analyses that present information stored in a container (e.g., PDF) as a set of components and their characteristics • Text – word frequency• Images (rasters) – color frequency (histogram)• Vector graphics – line frequency
• Exploration for appraisal analyses needs to include visible and invisible objectsvisible and invisible objects
Exploration of Text Components
LOADED FILESOccurrence of numbersOccurrence of words
“Ignore” words
Exploration of Image Components
Occurrence of colorsList of images Preview
LOADED FILES “Ignore” colors
Exploration of Vector Graphics ComponentsComponents
LOADED FILESLOADED FILESPreview Occurrence of v/h lines
Imaginations unbound
Exploration of Visible And Invisible Objects
Objects intersected at the mouse click location
Open Problem 2: Support of Appraisals by Enabling Comparisonsby Enabling Comparisons
• How to compare containers with heterogeneous i f ti ?information?• Methodology• Metrics• Weighting factors for fusion
• How to quantify differences between the same type of information? • Encodings and Representations• Metrics• Local versus global differences
Imaginations unbound
Comparisons
Imaginations unbound
MethodologyPartial solutions in literature +literature-Ref. CAPTCHA
+ …
Open problems
Relationship toPermanent Records
+ …
Experimental ExampleINPUT = 10 PDF docs (4 & 6 Groups)
UNIQUE ID= 1,2,3,4 UNIQUE ID= 5,6,7,8,9,10
Imaginations unbound
Comparative Experimental Results
INPUT = 10 PDF docs (6 & 4 members in each Group)(6 & 4 members in each Group)
V b d i il iVector-based similarity
Text-based similarity Image-based similarity
Comparative Experimental Results
Vector Graphics Similarity Portion of Document Surface and Word Similarity Combined Allotted to Each Document Feature
Comparison Using Combination of Document Features in Proportion to Coverage
Accuracy Comparisons
Method Average Similarity of
Average Similarity of
Average Similarity AcrossSimilarity of
Group 1Similarity of Group 2
Similarity Across Group 1 & 2
TEXT ONLY 1 0.489 0TEXT & IMAGE & 0 906 0 520 0 075TEXT & IMAGE & GRAPHICS
0.906 0.520 0.075
One refers to high similarity & zero refers to low similarity g y & y
Conclusions:•Differences in similarity are up to 10% of the score•Differences in similarity are up to 10% of the score•Documents in Group 2 would likely be misclassified as 0.5 similarity would be the threshold between similar and
Imaginations unbound
dissimilar documents
Open Problem 3: Scalability of Appraisals
• Scalability of appraisals with increasing heterogeneity of information, dimensionality of data and volume of electronic records
H h ld i l h• How should appraisal process change as 3D data is added to file containers?H h ld i l h• How should appraisal process change as 3D+time, 2D+spectrum, 3D+time+spectrum nD3D+time+spectrum, nD, …
• How should appraisal operations be designed to accommodate growingdesigned to accommodate growing volume of electronic records?
Imaginations unbound
Approaches to Computational Scalability of Document AppraisalsDocument Appraisals
• Options for parallel processing• message-passing interface (MPI)
MPI i d i d f h di i f i l i l• MPI is designed for the coordination of a program running as multiple processes in a distributed memory environment by using passing control messages.
• open multi processing (OpenMP)• open multi-processing (OpenMP)• OpenMP is intended for shared memory machines. It uses a
multithreading approach where the master threads forks any number of slave threadsnumber of slave threads.
• Map Reduce parallel programming paradigm for commodity clusters
It l t it i l M f ti d R d• It lets programmers write simple Map function and Reduce function, which are then automatically parallelized without requiring the programmers to code the details of parallel processes and communications
Imaginations unbound
processes and communications
• Specialized Hardware: FPGA, Cell processors, GPU
Computational Requirements forRequirements for Executing the MethodologyMethodology
Yellow indicatescomputations
Relationship toPermanent Records
Appraisal & Sampling
Hardware & Software Dependencies with HadoopHadoop• Test data: 15 PDF files from the Columbia investigation
web site at http://caib.nasa.gov/. p g• Software configuration: Linux OS (Ubuntu flavor) and
the Hadoop implementation of Map and Reduce f nctionalitiesfunctionalities
• Hardware configuration: homogeneous & heterogeneous machinesg
Hadoop Average Speed
405060
nds
0102030
1 2 3 4 5
#machines
seco
n average speed
Imaginations unbound
Homogeneous Hardware Heterogeneous Hardware
Open Problems Related to ArchivalOpen Problems Related to Archival Decisions
•Simulate Preservation Costs as a Function of Information•Simulate Preservation Costs as a Function of Information Granularity and Information Technology •Optimal Utilization of Computational and Human
Imaginations unbound
Resources
Open Problem: Archival Decision Support
• Decision support for forecasting preservation costs • How to predict computational and storage p p g
requirements of preservation as a function of technology variables and information gygranularity?
• How to optimize computational hardware,How to optimize computational hardware, software, storage, and networking investments?investments?
Imaginations unbound
Basic Questions About Information to be PreservedPreserved
National Center for Supercomputing Applications
Challenges in Forecasting• Volatility of software/hardware/storage media
• Updates: Windows operating systems since 2000: Two major new releases, two minor service pack updates, around fifty security , p p , y ypatches since SP2
• Upgrades: Microsoft Office Pro for Windows 95/98/ME/2000/XP/2003/2007
• Media life expectancy: Optical 5 years Disk 15 years Microfiche• Media life expectancy: Optical ~5 years, Disk ~ 15 years, Microfiche ~ 100, microfilm ~ 300, newspaper ~ 50, clay tablet ~ 10,000 (life expectancy vs. information density – [P. Conway, 1996] )
• Cost of software/hardware/storage media• Operating System: Windows 3.1/95/98/NT/2000/XP/Vista: Windows
95 = $209; Windows NT = $280; Windows XP = $300; Windows Vista = $399->$319 (2008)
• 128 MB of SDRAM: Year 1999 ~ $120-> $40 -> $200-250 due to128 MB of SDRAM: Year 1999 $120 > $40 > $200 250 due to Earthquake in Taiwan -> March 2000 ~ $55->March 2007 ~ $8.96 (flash card) - www.pricewatch.com (1TB ~$109.95 as of 01/15/2009)
• High performance computers: 2006: DARPA awards approximately $500 million to Cray and IBM; 2007 NSF $200 million to NCSA/IBM
National Center for Supercomputing Applications
$500 million to Cray and IBM; 2007 NSF $200 million to NCSA/IBM
Archival Decision Support
• Lack of forecasting models to predict preservation costs
• Our work: Understand the tradeoffs between information value and computational/storage costs by providing simulation frameworks• Information granularity, organization, compression, encryption,
document format, ...• Versus• Cost of CPU for gathering information, for processing and for
input/output operations; cost of storage media, upgrades, storage p p p ; g , pg , groom, …
• Prototype simulation framework: Image Provenance To Learn available for downloading fromLearn available for downloading from http://isda.ncsa.uiuc.edu
Simulation Framework
Decision Maker
Information Gathering and
Storage Learning
Information Retrieval and
Process Preservation
Value
Provenance Information
Reconstruction
Provenance Information
Preservation
Value
Information Information
Valu
e
observed
linear
Cost (memory, CPU)
Cost / Information Granularity Analysis
National Center for Supercomputing Applications
Image ViewerInformation Gathering System
Process Reconstruction System
Image Event Category Tracker
Events
Summary of Events
ViewedArea
Storage
Area
Storage
Time
Information Granularity
National Center for Supercomputing Applications
Storage vs. Information Organization Tradeoffs: Test Case
• Information granules include interpreted, raw and snapshots• Files were not compressed
Tradeoffs: Test Case
Files were not compressed
Event NameSaved Size
Mouse ClickedAdd Annotation
Change RGB BandChange Gray ScaleChange Auto Zoom
-RDF= Resource
Window ShownChange GammaWindow HiddenChange Selection
MagnificationMouse Clicked
RDF
Key Pair
Description Framework Metadata Model
1 10 100 1000 10000 100000 1000000 10000000
Window CreatedChange Zoom Factor
Change Visible RegionNew Image
-Key pair = XML Metadata Model
National Center for Supercomputing Applications
1 10 100 1000 10000 100000 1000000 10000000
Bytes (log scale)
Open Problems Related to AutomatingOpen Problems Related to Automating Archival Processing for Preservation
1. Discovery of Relationships Among Electronic Records2. Information Preserving Conversions of Electronic Records3. Sampling, Authenticity and Integrity Verification of a Collection
of Temporally Changing Records
Imaginations unbound
Open Problem 1: Discovering Relationships Among FilesRelationships Among Files
• How should one establish relationships among electronic records coming from disparate sources or from the samerecords coming from disparate sources or from the same source at multiple time instances?• How to extract metadata?• What ontology to use to represent the extracted
metadata?H t t t t d t t ti f lti l d t• How to automate metadata extraction from multiple data types, e.g., 2D drawings and 3D CAD models?
• How to discover relationships between electronic recordsHow to discover relationships between electronic records corresponding to the same physical objects but different multidimensional observations?
• Need to Understand the Complexity of the ProblemImaginations unbound
Metadata Extraction: Complexity & Size
the Crandon Mine Reports pfrom 1981 till 2003http://digicoll.library.wisc.edu/cgi-bin/EcoNatRes/EcoNatRes-idx?type=browse&scope=ECONATRES.CRANDONMINE
RDF t i l t t d i A t d i li d i RDF
Imaginations unbound
RDF triples extracted using Aperture and visualized using RDF-Gravity (red – edges, green-literal values, violet – properties)
Relationships Among Multiple Data Types• Example Data: Torpedo Weapon Retriever 841
• 784 existing 2D image drawings and N>22 3D CAD modelsmodels
• How to establish relationships among the 3D CAD models and 2D image drawings during a product lifecycle?
Hypothetical Distribution of 3D CAD models for
Imaginations unbound
Hypothetical Distribution of 3D CAD models for TWR 841
Understanding Challenges in Automation
ryD
isco
ver
nshi
p D
OCR Rel
atio
Descriptors (metadata)Representation
Imaginations unbound
Open Problem 2: Conversions of Electronic RecordsElectronic Records• Conversions of electronic records are needed because
• Visual exploration depends on various software packages
• Many formats are retired (deprecated) over timeA subset of formats is selected for preservation• A subset of formats is selected for preservation purposes
• How to measure the degree of information gpreservation when files are converted from format A to format B?• During conversions information could be lost added or modified• During conversions, information could be lost, added or modified• What is the importance of each byte, object, etc. ?
• How to introduce a framework for measuring the quality of conversion and visualization software?
Imaginations unbound
Example: Conversion of X3D to STEP to X3D
Software:
X3dToVrml97
WRLX3DSoftware:
A3D Reviewer
Software:
A3D ReviewerSoftware: Nothing!A3D ReviewerVrml97ToX3d
Nothing!
STEP WRL X3D
Automation of 3D File Format Mapping & ConversionConversion
Imaginations unbound
Open Problem 3: Sampling, Integrity and Authenticityg y y• Given finite resources and increasing amounts of electronic
records, automation of sampling, integrity and authenticity verification is very much neededverification is very much needed
• What are the criteria for sampling a collection of temporally changing versions of ‘the same’ document? • Authenticity• Integrity• Information content• Information content
• How to measure a degree of authenticity?• Computers might assign inaccurate time stamps to records
• How to detect integrity failures?• A record containing a female patient with prostate cancer
• How to incorporate constraints into sampling?• How to incorporate constraints into sampling?• Storage space, compression computational cost, etc.
Imaginations unbound
Example:Temporal Ranking and Integrity VerificationVerification
• Chronological ranking based on time stamps of filfiles• Last modification (current
implementation)
• Ranking can be changed by a human
• Content referring to• Content referring to dates can be used for integrity verification
Imaginations unbound
TIME
Rules and Attributes for Integrity Verification• Document integrity attributes?
• appearance or disappearance of document images• appearance and disappearance of dates embedded in
documents • file size • count of image groups• number of sentences• average value of dates found in a documentaverage value of dates found in a document
• Rules?
Imaginations unbound
Summary• Introduced a set of open problems
related toA i l f l t i d• Appraisal of electronic records
• Archival forecasting of preservation costscosts
• Automation of processing for preservationpreservation
• Examples used for illustrating the openExamples used for illustrating the open problems from our research just scratch the surface of some of the open
blproblems
Observations• Many stakeholders are already aware of some of the
open problems including government agencies and companies
• As all government agencies have been computerized, the continuity and functioning of the agencies depend on preservation and reconstruction of electronic records
• Right now, we are at the beginning of the exponential growth of electronic records (many more electronic records will be coming)
• Some scientific fields are already facing real time decisions about preserving electronic records (e.g.,
t )astronomers)
Future Vision
• It is envisioned that the preservation and reconstruction of electronic records have to follow different paradigms that incorporatefollow different paradigms that incorporate • Scalability (heterogeneity, dimensionality
and volume) )• Forecasting of preservation costs • New level of automation and quality
control in processing for preservationcontrol in processing for preservation purposes
• The field of electronic record managementThe field of electronic record management and preservation needs forward looking solutions to stay abreast with the dynamics y yof digital information
Imaginations unbound
References to Presented Research
• -Bajcsy P., R. Kooper and S-C. Lee, “Understanding Preservation and Reconstruction Requirements for Computer Assisted Decision Processes,” ACM Journal on Computers and Cultural Heritage (JOCCH), (submitted October 2008).
• -Bajcsy P., “A Perspective on Cyberinfrastructure for Water Research Driven by Informatics Methodologies,” GeographyBajcsy P., A Perspective on Cyberinfrastructure for Water Research Driven by Informatics Methodologies, Geography Compass, Volume 2, Issue 6 (p 2040-2061), 2008 Blackwell Publishing Ltd, URL: http://www3.interscience.wiley.com/cgi-bin/fulltext/121478978/PDFSTART
• -Bajcsy P., R. Kooper, L. Marini and J. Myers, “Community-Scale Cyberinfrastructure for Exploratory Science,” In: Cyberinfrastructure Technologies and Applications book, Editor: Junwei Cao, Nova Science Publishers, Chapter 12, Inc., 2009; URL: https://www.novapublishers.com/catalog/product info.php?products id=8011; p p g p _ p p p _
• - McHenry K. and P. Bajcsy "An Overview of 3D Data Content, File Formats and Viewers.", Technical Report NCSA-ISDA08-002, October 31, 2008
• -McFadden W., K. McHenry, R. Kooper, M. Ondrejcek, A. Yahja and P. Bajcsy, “Advanced Information Systems for Archival Appraisals of Contemporary Documents,” the 4th IEEE International Conference on e-Science, December 8-12, 2008, Indianapolis, IN., p ,
• -Lee S-C, W. McFadden and P. Bajcsy, “Text, Image and Vector Graphics Based Appraisal of Contemporary Documents,” The Seventh International Conference on Machine Learning and Applications, December 11-13, 2008, San Diego, CA.
• -Bajcsy P. and S-C Lee, "Computer Assisted Appraisal of Contemporary PDF Documents" ARCHIVES 2008: Archival R/Evolution & Identities 72nd Annual Meeting Pre-conference Programs: August 24-27, 2008, San Francisco, CA.& g g g , , ,
• -Lee S-C. and P. Bajcsy, “Understanding Challenges in Preserving and Reconstructing Computer-Assisted Medical Decision Processes,” the Workshop on Machine Learning in Biomedicine and Bioinformatics (MLBB07) of the 2007 International Conference on Machine Learning and Application (ICMLA07), Cincinnati, Ohio, December 13-15, 2007.
• -Bajcsy P and D. Clutter, “Gathering and Analyzing Information about Decision Making Processes Using Geospatial Electronic Records,” the 2006 Winter Federation of Earth Science Information Partners (“Federation”) Conference,Electronic Records, the 2006 Winter Federation of Earth Science Information Partners ( Federation ) Conference, poster, January 4-6, 2006 in Washington, DC.
Imaginations unbound
Questions
• Project URL: jhttp://isda.ncsa.uiuc.edu/NARA/index.htmland http://isda.ncsa.uiuc.edu/CompTradeoffs/
• Publications – see our URL at http://isda ncsa uiuc edu/publicationshttp://isda.ncsa.uiuc.edu/publications
• Peter Bajcsy; email: pbajcsy@ncsa uiuc edu• Peter Bajcsy; email: [email protected]