@string{aaai = "Proceedings AAAI"}
@string{ai   = "Artificial Intelligence"}
@string{berk = "University of California Berkeley"}
@string{biocyb = "Biological Cybernetics"}
@string{cmu = "Carnegie-Mellon University"}
@string{cacm = "Communications of the ACM"}
@string{capami = "Proc. 1987 Workshop on Comp. Arch. for Patt. Anal. and Mach. Intell."}
@string{computer = "IEEE Computer"}
@string{cgip = "Computer Graphics and Image Processing"}
@string{cvgip = "Computer Vision, Graphics, and Image Processing"}
@string{cvpr = "Proceedings IEEE Conf. on Computer Vision and Pattern Recognition"}
@string{focs = "Proceedings Symposium on Foundations of Comp. Sci."}
@string{iccv = "Proceedings of the International Conference on Computer Vision"}
@string{icpp = "Proceedings Int. Conf. on Parallel Processing"}
@string{icpr = "Proceedings Int. Conf. on Pattern Recognition"}
@string{ieee = "Proceedings  of the IEEE"}
@string{ijcai = "Proceedings IJCAI"}
@string{ijcv = "International Journal of Computer Vision"}
@string{iu = "Proceedings Image Understanding Workshop"}
@string{josa = "Journal of the Optical Society of America"}
@string{jrr = "International Journal of Robotics Research"}
@string{mit = "Massachusetts Institute of Technology"}
@string{mitai = "Artificial Intelligence Laboratory, Massachusetts Institute of Technology"}
@string{opteng = "Optical Engineering"}
@string{pami = "IEEE Transactions on Pattern Analysis and Machine Intelligence"}
@string{patt = "Proceedings Int. Conf. on Pattern Recognition"}
@string{pnas = "Proceedings of the National Academy of Science"}
@string{pr = "Pattern Recognition"}
@string{prip = "Proceedings of IEEE Computer Society Conference on Pattern Recognition and Image Processing"}
@string{prsl = "Proceedings of the  Royal Society of London"}
@string{prslb = "Proceedings of the  Royal Society of London B"}
@string{robau = "Proceedings of IEEE Conference on Robotics and Automation"}
@string{sciam = "Scientific American"}
@string{siam = "SIAM J. Comp."}
@string{stoc = "Proceedings ACM Symposium on Theory of Computing"}
@string{tmc = "Thinking Machines Corporation"}
@string{visres = "Vision Research"}
@string{jcogneuro = "Journal of Cognitive Neuroscience"}
@string{jexppsych = "Journal of Experimental Psychology"}
@string{annrevneuro= "Annual Review of Neuroscience"}

@string{Ailab = "MIT Artificial Intelligence Laboratory"}
@string{media-vision = "MIT Media Laboratory Vision and Modeling Group"}
@string{aaai-87 = "Sixth National Conference on Artificial Intelligence"}
@string{bbs = "Brain and Behavioral Sciences"}
@string{belknap = "The Belknap Press of Harvard University Press"}
@string{mit = "Massachusetts Institute of Technology"}
@string{mitpress = "MIT Press"}
@string{ieee-proc = "Proceedings of the IEEE"}
@string{ieee-ra = "IEEE Journal of Robotics and Automation"}
@string{ieee-ra-86 = "The 1986 IEEE Conference an Robotics and Automation"}
@string{ieee-smc = "IEEE Transactions on Systems, Man, and Cybernetics"}
@string{ieee-com = "IEEE Transactions on Communications"}
@string{ieee-nn = "IEEE Transactions on Neural Networks"}
@string{ijcai-87 = "Tenth International Joint Conference on Artificial Intelligence"}
@string{umass = "University of Massachusetts at Amherst"}
@string{Tnips2 = "Advances in Neural Information Processing 2"}
@string{Tnips3 = "Advances in Neural Information Processing 3"}
@string{Tnips4 = "Advances in Neural Information Processing 4"}
@string{Tnips5 = "Advances in Neural Information Processing 5"}
@string{Tnips6 = "Advances in Neural Information Processing 6"}
@String{jneusci= "Journal of Neuroscience"}
@string{icsi = "Internation Computer Science Institute"}
@string{neuropros = "Placed in the Neuroprose archive."}
@string{aaaisymp93 = "AAAI Fall Symposium Series Working Notes"}
@string{colt = "Proceedings of the Conference on Computational Learning Theory"}
@string{ACM = "Association for Computing Machinery"}


@proceedings{NN:nips2,
     editor = "David S. Touretzky",
      title = nips,
  booktitle = nips,
     volume = 2,
    address = "Denver 1989",
  publisher = "Morgan Kaufmann, San Mateo",
       year = 1990
}

@proceedings{NN:nips3,
     editor = "Richard P. Lippmann and John E. Moody and David S. Touretzky",
      title = nips,
  booktitle = nips,
     volume = 3,
    address = "Denver 1990",
  publisher = "Morgan Kaufmann, San Mateo",
       year = 1991
}

@proceedings{NN:nips4,
     editor = "John E. Moody and Steven J. Hanson and Richard P. Lippmann",
      title = nips,
  booktitle = nips,
     volume = 4,
    address = "Denver 1991",
  publisher = "Morgan Kaufmann, San Mateo",
       year = 1992
}

@proceedings{NN:nips5,
     editor = "Steven J. Hanson and Jack D. Cowan and C. Lee Giles",
      title = nips,
  booktitle = nips,
     volume = 5,
    address = "Denver 1992",
  publisher = "Morgan Kaufmann, San Mateo",
       year = "1993"
}

@proceedings{NN:nips6,
     editor = "Jack D. Cowan and Gerald Tesauro and Joshua Alspector",
      title = nips,
  booktitle = nips,
     volume = 6,
    address = "Denver 1993",
  publisher = "Morgan Kaufmann, San Francisco",
       year = "1994"
}
@proceedings{NN:nips7,
     editor = "",
      title = nips,
  booktitle = nips,
     volume = 7,
    address = "Denver 1994",
  publisher = "Morgan Kaufmann, San Francisco",
       year = "1995"
}


@proceedings{NN:ijcnn89,
      title = ijcnn,
  booktitle = ijcnn,
  publisher = "IEEE",
       year = 1989
}

@proceedings{NN:ijcnn92,
      title = ijcnn,
  booktitle = ijcnn,
    address = "Baltimore 1992",
  publisher = "IEEE",
       year = 1992
}

%%%%%%%%%%%%%%%%%%%%%% end cross references %%%%%%%%%%%%%%%%%%%%%%

@MISC{ferrel90,
	AUTHOR = "Cynthia Ferrel",
	TITLE = {Description of Thesis Topic},
	HOWPUBLISHED = {Personal Communication},
	YEAR = 1990
}

@MISC{horswill90,
	AUTHOR = {Ian Horswill},
	TITLE = {Thesis Proposal},
	HOWPUBLISHED = {In Preparation},
	YEAR = {1990}
}

@PROCEEDINGS{widrow60,
	AUTHOR = {Bernard Widrow and Marcian E. Hoff},
	TITLE = {Adaptive Switching Circuits},
	YEAR = {1960},
	ORGANIZATION = {IRE WESCON Convection Record},
}

@BOOK{arbib87,
	AUTHOR = "Michael A. Arbib",
	TITLE = "Brains, Machines, and Mathematics",
	PUBLISHER = "Springer Verlag",
	YEAR = 1987,
	ADDRESS = "New York, N.Y.",
	EDITION = "Second"
}

@ARTICLE{agre90,
	AUTHOR = "Philip E. Agre",
	TITLE = "Agency and Structure: Ten Issues for Computational Research",
	JOURNAL = AAAI,
	YEAR = 1990,
	ANNOTE = "Interesting overview of where AI and artificial creatures
	should be going in the future."
}

@BOOK{ogata70,
	AUTHOR = "Katsuhiko Ogata",
	TITLE = "Modern Control Engineering",
	PUBLISHER = "Prentice-Hall",
	YEAR = 1970,
	ADDRESS = "Englewood Cliffs, N.J.",
	ANNOTE = "Steady State Frequency Response (page 372)"
}

@ARTICLE{gonshor76,
	AUTHOR = "A. Gonshor and M. Jones ",
	TITLE = "Extreme vestibulo-ocular adaptation induced by prolonged optical reversal of vision",
	JOURNAL = "J. Physiol.",
	YEAR = 1976,
	NUMBER = 256,
	PAGES = "381-414",
	ANNOTE = "Haven't read this yet."
}



@ARTICLE{miles81,
	AUTHOR = "F. A. Miles and S. G. Lisberger",
	TITLE = "Plasticity in the Vestibulo-ocular Reflex: A New Hypothesis",
	JOURNAL = "Ann. Rev. Neurosci.",
	YEAR = 1981,
	VOLUME = 4,
	PAGES = "273-299"
}



@ARTICLE{lisberger88,
	AUTHOR = "Stephen G. Lisberger",
	TITLE = "The Neural Basis for Learning of Simple Motor Skills",
	JOURNAL = "Science",
	YEAR = 1988,
	VOLUME = 242,
	PAGES = "728-735"
}



@ARTICLE{moody89,
	AUTHOR = "John Moody and Christian J. Darken",
	TITLE = "Fast Learning in Networks of locally-tuned Processing
Units",
	JOURNAL = "Neural Computation",
	YEAR = 1989,
	VOLUME = 1,
	PAGES = "281-294"
}



@ARTICLE{miles87,
	AUTHOR = "F.A. Miles and K. Kawano",
	TITLE = "Visual Stabilization of the Eyes",
	JOURNAL = "TINS",
	YEAR = 1987,
	VOLUME = 4,
	NUMBER = 10,
	PAGES = "153-158",
	ANNOTE = "Reference on Opto-kinetic nystagmus latency"
}

@ARTICLE{poggio73,
	AUTHOR = "T. Poggio and W. Reichard",
	TITLE = "Considerations on Models of Movement Detection",
	JOURNAL = "Kybernetic",
	YEAR = 1973,
	VOLUME = 13,
	PAGES = "223-227"
}

@BOOK{hildreth83,
	AUTHOR = "Ellen C. Hildreth",
	TITLE = "The Measurement of Visual Motion",
	PUBLISHER = "The MIT Press",
	YEAR = 1983,
	ANNOTE = "Good book on the extraction of motion from edges."
}

@INCOLLECTION{gouras85,
	AUTHOR = "Peter Gouras",
	TITLE = "Oculomotor System",
	BOOKTITLE = "Principles of Neuroscience",
	PUBLISHER = "Elsevier Science Publishing",
	YEAR = 1985,
	EDITOR = "Eric Kandel and James Schwartz",
	CHAPTER = 34
}



@BOOK{kandel85,
	EDITOR = "Eric Kandel and James Schwartz",
	TITLE = "Principles of Neuroscience",
	PUBLISHER = "Elsevier Science Publishing",
	YEAR = 1985
}

@TECHREPORT{brooks89a,
	AUTHOR = "Rodney A. Brooks, Pattie Maes, Maja J. Mataric, Grinell More",
	TITLE = "Lunar Base Construction Robots",
	INSTITUTION = "MIT",
	YEAR = 1989,
	TYPE = "AI Memo",
	NUMBER = "???",
	MONTH = "November",
	ANNOTE = "Roving simple autonmous moon robots.  Complexity of
behavior comes from a reliance on emergent behavior.  While it seems
that each robot will be simpler than one large robot that would
perform the task, the interactional phenomena may be very difficult to
analyze.  I have yet to see a convincing argument for why this stuff
should be manifestly more simple than anything else.  Filed under
Pattie Maes."
}

@ARTICLE{raibert89,
	AUTHOR = "Marc H. Raibert",
	TITLE = "Trotting, Pacing, and Bounding by a Quadruped Robot",
	JOURNAL = "Journal of Biomechanics",
	YEAR = 1989
}

@MASTERSTHESIS{angle89,
	AUTHOR = "Colin Angle",
	TITLE = "Genghis, a Six Legged Autonomous Walking Robot",
	SCHOOL = "MIT",
	YEAR = "1989"
}

@TECHREPORT{omohundro87,
	AUTHOR = "Stephen M. Omohundro",
	TITLE = "Efficient Algorithms with Neural Network Behavior",
	INSTITUTION = "University of Illinois",
	YEAR = 1987,
	NUMBER = 1331,
	ANNOTE = "Good serial implemtations of nn algs."
}



@ARTICLE{lippmann87,
	AUTHOR = "Richard P. Lippmann",
	TITLE = "An Introduction to Computing with Neural Nets",
	JOURNAL = "IEEE ASSP Magazine",
	YEAR = 1987,
	MONTH = "April",
	ANNOTE = "Good overview of the Neural network lit."
}

@TECHREPORT{hinton87,
	AUTHOR = "Geoffrey E. Hinton",
	TITLE = "Connectionist Learning Procedures",
	INSTITUTION = "Artificial Intelligence",
	YEAR = 1987,
	TYPE = "CS",
	NUMBER = "87-115",
	ANNOTE = "Good overview of the Neural network lit."
}

@ARTICLE{eckmiller83,
	AUTHOR = "Rolf Eckmiller",
	TITLE = "Neural Control of Foveal Pursuit Versus Saccadic Eye Movements in Primates --- Single Unit Data and Models",
	JOURNAL = "IEEE Transactions on Systems, Man, and Cybernetics",
	YEAR = 1983,
	VOLUME = 13,
	NUMBER = 5,
	ANNOTE = "Haven't read it yet."
}

@INPROCEEDINGS{grosse88,
	AUTHOR = "Eric Grosse",
	TITLE = "A Catalog of Algorithms for Approximation",
	BOOKTITLE = "Proceedings of the Algorithms for Approximation Conference",
	YEAR = 1988,
	PUBLISHER = "Chapman and Hall",
	ANNOTE = "An incredible bibliography on approximation"
}

@INPROCEEDINGS{Barron88,
	AUTHOR = "A. Barron and R. Barron",
	TITLE = "Satatistical Learning Networks: A Unifying View",
	BOOKTITLE = "Computing Science and Statistics",
	YEAR = 1988,
	ANNOTE = "Got this from CGA need to get a better reference."
}

@INCOLLECTION{rumelhart86b,
	AUTHOR = "David E. Rumelhart and Geoffrey E. Hinton and R. J. Williams",
	TITLE = "Learning internal representations by error propagation",
	BOOKTITLE = "Parallel Distributed Processing: Exploration in the Microstructure
of Cognition",
	PUBLISHER = "MIT Press",
	YEAR = 1986,
	CHAPTER = 8,
	EDITOR = "David E. Rumelhart and James L. McLelland"
}

@BOOK{rumelhart86a,
	EDITOR = "David E. Rumelhart and James L. McLelland",
	TITLE = "Parallel Distributed Processing: Exploration in the
Microstructure of Cognition",
	PUBLISHER = "MIT Press",
	YEAR = 1986
}

@article{PogTorKoc85,
author = "Tomaso Poggio and Vincent Torre and Christof Koch",
title = "Computational Vision and Regularization Theory",
journal = "Nature",
volume = 317,
year = 1985,
pages = "314--319"
}

@BOOK{press88,
	AUTHOR = "W. Press et al",
	TITLE = "Numerial Recipes in C: The Art of Scientific Programming",
	PUBLISHER = "Cambridge University Press",
	YEAR = 1988,
	ANNOTE = "Great book full of c programs"
}

@BOOK{moravec80,
	AUTHOR = "Hans P. Moravec",
	TITLE = "Robot Rover Navigation",
	PUBLISHER = "UMI Research Press",
	YEAR = 1980
}

@TECHREPORT{connell88,
	AUTHOR = "Jonathan H. Connel",
	TITLE = "A Behavior-Based Arm Controller",
	INSTITUTION = "MIT",
	YEAR = 1988,
	TYPE = "TR",
	NUMBER = 1025
}

@TECHREPORT{connell89,
	AUTHOR = "Jonathan H. Connel",
	TITLE = "A Colony Architecture for and Artificial Creature",
	INSTITUTION = "MIT",
	YEAR = 1989,
	TYPE = "TR",
	NUMBER = 1151
}

@ARTICLE{robinson81,
	AUTHOR = "D. A. Robinson",
	TITLE = "The Use of Contorl Systems in the Neurophysiology of Eye Movements",
	JOURNAL = "Ann. Rev. Neurosciencd",
	YEAR = 1981,
	VOLUME = 4,
	PAGES = "463-503"
}


@INCOLLECTION{robinson85,
	AUTHOR = "D. A. Robinson",
	TITLE = "The Contorl of Eye Movements",
	BOOKTITLE = "Handbook of Physiology -- The Nervous System II",
	YEAR = 1985,
	CHAPTER = 28,
	PUBLISHER = "??"
}



@ARTICLE{collewijn72,
	AUTHOR = "H. Collewijn",
	TITLE = "An Analog Model of the Rabbits Optokinetic System",
	JOURNAL = "Brain Reasearch",
	YEAR = 1972,
	VOLUME = 36,
	PAGES = "71-88"
}

@BOOK{yarbus67,
	AUTHOR = "A. L. Yarbus",
	TITLE = "Eye Movements and Vision",
	PUBLISHER = "Plenum Press",
	YEAR = 1967
}

@BOOK{rogers76,
	AUTHOR = "David F. Rogers and J. Alan Adams",
	TITLE = "Mathematical Elements for Computer Graphics",
	PUBLISHER = "McGraw-Hill",
	YEAR = 1976

}

@CONFERENCE{burt89,
	AUTHOR = "P. J. Burt et al",
	TITLE = "Object Tracking with a Moving Camera",
	BOOKTITLE = "IEEE Workshop on Motion",
	YEAR = 1989,
	ORGANIZATION = "IEEE",
	ADDRESS = "Irvine, CA"
}

@MASTERSTHESIS{cornog82,
	AUTHOR = "Katherine H. Cornog",
	TITLE = "Smooth Pursuit and Fixation for Robot Vision",
	SCHOOL = "MIT",
	YEAR = 1982
}

@PHDTHESIS{mel89,
	AUTHOR = "Bartlett W. Mel",
	TITLE = "MURPHY: A Neurally-Inspired Connectionist Approach to
Learning and Performance in Vision-Based Robot Motion Planning",
	SCHOOL = "University of Illinois",
	YEAR = 1989
}

@CONFERENCE{thompson87,
	AUTHOR = "William B. Thompson and Ting-Chuen Pong",
	TITLE = "Detecting Moving Objects",
	BOOKTITLE = "First International Conference on Computer Vision",
	YEAR = 1987,
	ORGANIZATION = "IEEE"
}

@ARTICLE{miller87a,
	AUTHOR = "W. Thomas Miller IIIet al",
	TITLE = "Application of a General Learning Algorithm to the
Control of Robotic Manipulators",
	JOURNAL = "The International Journal of Robotics Reasearch",
	YEAR = 1987,
	VOLUME = 6,
	NUMBER = 2
}

@ARTICLE{miller87b,
	AUTHOR = "W. Thomas Miller III",
	TITLE = "Sensor-Baed Control of Robotic Manipulators Using a
General Learning Algorithm",
	JOURNAL = "IEEE Journal of Robotics and Automation",
	YEAR = 1987,
	VOLUME = "RA-3",
	NUMBER = 2
}

@CONFERENCE{atkeson90,
	AUTHOR = "Christopher G. Atkeson",
	TITLE = "Using Local Models to Control Movement",
	BOOKTITLE = "Neural Information Processing 2",
	YEAR = 1989,
	EDITOR = "David S. Touretzky",
	PAGES = "316-324",
	PUBLISHER = "Morgan-Kaufmann"
}

@BOOK{atkeson88,
	AUTHOR = "Chae H. An and Cris G. Atkeson and John M. Hollerbach",
	TITLE = "Model-Based Control of a Robot Manipulator",
	PUBLISHER = "MIT Press",
	YEAR = 1988
}

@ARTICLE{clark?,
	AUTHOR = "James J. Clark and Nicola J. Ferrier",
	TITLE = "Modal Control of an Attentive Vision System",
	JOURNAL = "????",
	YEAR = "???"
}

@ARTICLE{moravec83,
	AUTHOR = "Hans P. Moravec",
	TITLE = "The Stanford Cart and the CMU Rover",
	JOURNAL = "Proceedings of the IEEE",
	YEAR = 1983,
	VOLUME = 71,
	NUMBER = 7,
	PAGES = "872-884",
	ANNOTE="Overview of the work at CMU and Stanford"
}

@ARTICLE{turk88,
	AUTHOR = "M. A. Turk et al",
	TITLE = "VITS - A Vision System for Autonomous Land Vehicle
Navigation",
	JOURNAL = pami,
	YEAR = 1987,
	VOLUME = 10,
	NUMBER = 3,
	PAGES = "342-361"
}

@ARTICLE{thorpe88,
	AUTHOR = "Chuck Thorpe et al",
	TITLE = "Vision and Navigation for the Carnegie-Mellon Navlab",
	JOURNAL = pami,
	YEAR = 1988,
	VOLUME = 10,
	NUMBER = 3
}

@article{PogRei76,
author = "Tomaso Poggio and Werner Reichardt",
title = "Visual Control of Orientation Behaviour in the Fly: Part {II}: Towards the Underlying Neural Interactions",
journal = "Quart. Rev. Biophysics",
volume = 9,
year = 1976,
pages = "377--438"
}




@INCOLLECTION{deubel87,
	AUTHOR = "Heiner Deubel",
	TITLE = "Adaptivity of Gain and Direction in Oblique Saccades",
	BOOKTITLE = "Eye Movements",
	PUBLISHER = "Elsevier Science Publishers",
	YEAR = 1987,
	EDITOR = "J. K. O'Regan and A. Levy-Schoen",
	PAGES = "181-190"
}

@ARTICLE{barnes79,
	AUTHOR = "G. R. Barnes",
	TITLE = "Vestibulo-ocular Function During Co-ordinated Head
and Eye Movements to Acquire Visual Targets",
	JOURNAL = "J. Physiol.",
	YEAR = 1979
}

@ARTICLE{rashbass61,
	AUTHOR = "C. Rashbass",
	TITLE = "The Relationship Between Sasccadic and Smooth
Tracking Eye Movements",
	JOURNAL = "J. Physiol.",
	YEAR = 1961,
	VOLUME = 159,
	PAGES = "326-338"
}

@ARTICLE{yasui84,
	AUTHOR = "S. Yasui and L. R. Young",
	TITLE = "On the Predictive Control of Foveal Eye Tracking and
Slow Phases of Optokinetic and Vestibular Nystagmus",
	JOURNAL = "J. Pysiol.",
	YEAR = 1984,
	VOLUME = 347,
	PAGES = "17-33"
}

@ARTICLE{collewijn84,
	AUTHOR = "Han Collewijn and Ernst P. Tamminga",
	TITLE = "Human Smooth and Saccadic Eye Movements During
Voluntary Pursuit of Different Target Motions on Different Backgrounds",
	JOURNAL = "J. Physiol.",
	YEAR = 1984,
	VOLUME = 351,
	PAGES = "217-250",
	ANNOTE = "Full description of tracking in a variety of
situations including sinusoidal motion and different backgrounds"
}

@article{TorPog78,
author = "Vincent Torre and Tomaso Poggio",
title = "A Synaptic Mechanism Possibly Underlying Directional Selectivity to Motion",
journal = "Proc. R. Soc. Lond. B",
volume = 202,
pages = "409--416",
year = 1978}

@MASTERSTHESIS{horswill88,
	AUTHOR = "Ian D. Horswill",
	TITLE = "Reactive Navigation for Mobile Robots",
	SCHOOL = "MIT",
	YEAR = 1988
}

@ARTICLE{angle90,
	AUTHOR = "Colin Angle and Rodney Brooks",
	TITLE = "Small Planetary Rovers",
	JOURNAL = "IEEE Int. Workshop on Intell. Robots and Sensors",
	YEAR = 1990
}

@ARTICLE{ballard89,
	AUTHOR = "Dana H. Ballard",
	TITLE = "Reference Frames for Animate Vision",
	JOURNAL = "International Joint Conference on Artificial
Intelligence",
	YEAR = 1989
}

@ARTICLE{brown90,
	AUTHOR = "Christopher Brown",
	TITLE = "Gaze Contols with Interactions and Delays",
	JOURNAL = "IEEE Transactions on Systems, Man, and Cybernetics",
	YEAR = 1990,
	VOLUME = 20,
	NUMBER = 1,
	PAGES = "518-527"
}

@book{marr82,
author = "David Marr",
title = "Vision: A Computational Investigation into the Human Representation and Processing of Visual Information",
address = "San Francisco",
publisher = "W.H. Freeman and Company",
year = 1982}


@article{ullman81,
author = "Shimon Ullman",
title = "Analysis of Visual Motion by Biological and Computer Systems",
journal = computer,
month =  "August",
year = 1981,
volume = 14,
pages = "57--69"
}

@ARTICLE{aloimonos87,
	AUTHOR = "J. Aloimonos, I. Weiss and A. Bandopadhay",
	TITLE = "Active Vision",
	JOURNAL = iccv,
	YEAR = 1987,
	PAGES = "35-54"
}



@InBook{lenat83,
	Author= "D. B. Lenat",
	BookTitle= "Machine Learning",
        editor= "R. S. Michalski and J. G. Carbonell and T. M. Mitchell",
	Title= "The Role of Heuristics in Learning By Discovery: Three Case Studies",
	chapter= 9,
       publisher= "Tioga Publishing Company",
       volume= 1,
       year= 1983
	}

@article{brooks86a,
	author= "Rodney A. Brooks",
	year= 1986,
	month=apr,
	title= "A Robust Layered Control System for a Mobile Robot",
	journal=ieee-ra,
	volume= "RA-2",
	number= 1}

@inproceedings{brooks86b,
       author= "Rodney A. Brooks an Jonathan H. Connell",
       year= 1986,
       month=oct,
       title= "Asynchronous Distributed Control System for a Mobile Robot",
       booktitle= "SPIE's Cambridge Symposium on Optical and Optoelectrical Engineering"} 

@book{holland75,
	title= "Adaptation in Natural and Artificial Systems",
	author= "J. H. Holland",
	year= 1975,
	publisher= "The Michigan University Press"}

@article {hinton87b,
       author= "G. E. Hinton and S. J. Nowlan",
       title= "How Learning Can Guide Evolution",
       journal= "Complex Systems",
       volume= 1,
       year= 1987}

@inproceedings{brooks86c,
	author= "Rodney A. Brooks and Jonathan H. Connell and Anita Flynn",
	year= 1986,
	month=aug,
	title= "A Mobile Robot with Onboard Parallel Processor and
		Large Workspace Arm",
	booktitle= "AAAI-86"}

@inproceedings{connell87,
   author= "J. H. Connell",
   title= "Creature Design with the Subsumption Architecture",
   booktitle= "IJCAI",
   year= 1987}

@techreport{brooks88,
       author= "Rodney A. Brooks and Jonathan H. Connell and Peter Ning",
       title= "Herbert: A Second Generation Mobile Robot",
	Year= 1988,
	Month=jan,
	institution=Ailab,
	Number= 1016,
	Type= "A.I. Memo"
	}

@book{feller68,
       author= "W. Feller",
       year= 1968,
       title= "An Introduction to Probability Theory and Its Applications",
       volume= 1,
       publisher= "John Wiley and Sons",
       edition= "Third"}


@BOOK{strang86,
	AUTHOR = {Gilbert Strang},
	TITLE = {Introduction to Applied Mathematics},
	PUBLISHER = {Wellesley-Cambridge Press},
	YEAR = {1986}
}

@ARTICLE{newsome88,
	AUTHOR = "W.T. Newsome and E. B. Pare",
	TITLE = "A selective impairment of motion perception following lesions of the middle temporal visual area ({MT})",
	JOURNAL = "J. Neuroscience",
	YEAR = 1988,
	VOLUME = 8,
	PAGES = "2201-2211"
}

@ARTICLE{krauzlis89,
	AUTHOR = "R.J. Krauzlis and S. G. Lisberger",
	TITLE = "A control systems model of smooth pursuit eye movements with realistic emergent properties",
	JOURNAL = "Neural Computation",
	YEAR = 1992,
	VOLUME = 1,
	PAGES = "116-122"
}

@ARTICLE{ballard91,
	AUTHOR = "D. H. Ballard",
	TITLE = "Animate vision",
	JOURNAL = "Artificial Intelligence",
	YEAR = 1991,
	VOLUME = 48,
	PAGES = "57-86",
}

@UNPUBLISHED{lisberger92,
	AUTHOR = "S.G. Lisberger and T.J. Sejnowski",
	TITLE = "Computational analysis suggests a new hypothesis for motor learning  in the vestibulo-ocular reflex",
	NOTE = "Submitted for publication.",
	YEAR = 1992
}

@ARTICLE{adelson85,
	AUTHOR = "E. H. Adelson and J. R. Bergen",
	TITLE = "Spatiotemporal energy models of the perception of motion",
	JOURNAL = josa,
	YEAR = 1985,
	VOLUME = 2,
	NUMBER = 2,
	PAGES = "284-299"
}

@INCOLLECTION{robinson71,
	AUTHOR = "D. A. Robinson",
	TITLE = "Models of oculomotor neural organization",
	BOOKTITLE = "The Control of Eye Movements",
	YEAR = 1971,
	EDITOR = "P. Bach-y-Rita and C. C. Collins",
	PAGES = "519",
	ADDRESS = "New York",
	PUBLISHER = "Academic"
}

@ARTICLE{lisberger87,
	AUTHOR = "S. G. Lisberger and E. J. Morris and L. Tychsen",
	TITLE = "",
	JOURNAL = "Ann. Rev. Neurosci.",
	YEAR = 1987,
	VOLUME = 10,
	PAGES = "97-129",
}

@Article{fukushima89,
  author = 	 "Kunihiko Fukushima",
  title = 	 "Analysis of the Process of Visual Pattern
		  Recognition by the Neogcognitron",
  journal = 	 "Neural Networks",
  year = 	 1989,
  volume = 	 2,
  pages = 	 "413-420",
  file=       "6-5-93",

  annote = "An overview of the Neo.  Multi-level arch.  Each level
extracts a number of different features from the previous level --
using a form of weight sharing the same features are extracted
everywhere.  The next level collects information across a local region
about the same feature -- essentially blurring the position of the
feature.  Subsequent levels uses the features from the previous level
as their input.

There is an unsupervised training scheme proposed that is somewhat
interesting.  It is basically an associative mechanism.  It is unclear
if it was ever really tested.  "

}


@Article{spirkovska93,
  author = 	 "Lilly Spirkovska and Max B. Reid",
  title = 	 "Coarse-Coded Higher-Order Neural Networks for PSRI
		  Object Recognition",
  journal = 	 ieee-nn,
  year = 	 1993,
  volume = 	 4,
  number = 	 2,
  pages = 	 "276-283",
  month = 	 "March",
  file=       "6-5-93",

  OPTannote = "Phew.  This is really a stupid one.  They want to take
advantage of the NN craze but procede to propose a computationally
intractible approach to vision.  First they use binary outlines as
their inputs.  Then they construct a network that has a multiplicative
hidden unit for every triplet of pixels --- since the inputs are
binary the hidden layer computes which of the possible triangles in
the input are present.  Then they constrain the weights from the hidden
units whose triplets come from a triangle that is similar be the
same.  Since this is done at the pixel level there is a threshold for
similarity, all triangles that are almost the same have the same
weight.  Then they train the output layer to recognize the objects.  

So what they are really doing is computing all possible pixel
triangles.  Grouping them into clusters based on approximate
similarity.  Computing a representation vector based on the number of
triangles in each cluster.   Learning in this case is unecessary,
nearest neighbor would work just fine.

They insist on making a big deal out of the position, rotation and
scale invariance that they get.  But clearly triangle similarity is
invariant to these transformations.  

Their paper is supposedly about the use of coarse coding to reduce the
ridiculous number of hidden units -- triangles, computed.  They use a
number of overlapping coarse grids instead of a single grid.  This
keeps down the number of triangles generated and makes the triangle
buckets larger.  Low and behold on a two image test set,
classification is perfect.  But there is no analysis on how this might
work with a larger number of examples." 

}

@Article{burns93,
  author = 	 "John A. Burns, Paul A. Viola and George M. Whitesides",
  title = 	 "Predicting Musk Odor with Feed-Forward Neural Networks",
  journal = 	 "???",
  year = 	 1993,
  file=		 "6-5-93",
  annote =	 "Paper written by John on the work I did at Arris on
predicting Musk odor.  It hasn't been accepted yet."

}


@Article{chelazzi93,
  author = 	 "Leonardo Chelazzi, Earl K. Miller, John Duncan and
                 Robert Desimone",
  title = 	 "A Neural Basis for Visual Search in Inferior
                 Temporal Cortex",
  journal = 	 "Nature",
  year = 	 1993,
  volume = 	 363,
  pages = 	 "345-347",
  month = 	 "May",
  file=		 "6-9-93",

  OPTannote = "Very interesting.  

Authors wanted to know whether attention might effect the response of
a cell to stimuli.  A cell is identified that responds well to one
stimulus and poorly to another.  The monkey is forced to choose
between the two when presented together -- in the receptive field of
the cell.  The monkey chooses the one that it displayed right before a
delay.  In every case the choice display is the same: the weak and the
strong stimulus. If the monkey is shown the strong stimulus initially,
the responses are as you might expect -- strong response during the
initial presentation, and strong during the choice.  But if the monkey
must choose the weak stimulus, the response of the cell seems to be
captured.  The response at the beginning is weak.  But, then during
the choice the response is initially strong, and then is clamped down
to normal levels.

So even though the choice display is the same, the response is very
different depending on what the monkey is looking for.  Annote the
positions of the objects in the choice displays is randomized.  The
position of the target and distractor is not predictable.  Previous
work describes location dependant attentional phenomena.  This is the
first work that describes task dependant, or feature dependant,
attention.

" }

@Article{cowey93,
  author = 	 "Alan Cowey",
  title = 	 "Seeing the Tree for the Woods -- News and Views",
  journal = 	 "Nature",
  year = 	 1993,
  volume = 	 363,
  pages = 	 298,
  month = 	 "May",
  file=		 "6-9-93",
  annote =	 "This refers to chelazzi93."
}

@Article{MorDes85,
  author = 	 "Jeffrey Moran and Robert Desimone",
  title = 	 "Selective Attention Gates Visual Processing in the
                  Extrastiate Cortex",
  journal = 	 "Science",
  year = 	 1985,
  volume = 	 229,
  pages = 	 "782-784",
  month =	 "August",
  file=		 "6-9-93",

  annote =	 "The response of a V4 cell (which has a small 2-4 degree
receptive field) is dependant on the point of attention in of a
monkey.  A monkey is asked to respond in a match to sample at a
particular location.  The response of a V4 cell that has a receptive
field that includes the attended point is captured.  The response to a
stimuli that makes the cell respond strongly is effected by this
attention.  Response is attenuated if it is presented away from the
attended area.  Response is not attenuated if the strong stimuli is
presented in the attended area.  When the monkey is asked to attend to
an area outside of the receptive field the cell's response is not
attenuated.

In IT cells, where the receptive field is quite large, activity shows
a similar attenuation.  It was not technically possible to place the
attended area outside of the receptive field.  

In V1, where the receptive fields are less than 1 degree, there were
no measured effect of attention."

}

@InCollection{bridle89,
  author = 	 "John S. Bridle",
  title = 	 "Training Stochastic Model Recognition Algorithms as
		  Networks can lead to Maximum Mutual Information
		  Estimation of Parameters",
  booktitle = 	 Tnips2,
  publisher = 	 "Morgan Kaufman",
  year = 	 "1989",
  editor = 	 "David S. Touretzky",
  pages = 	 "211-217",
}


@InCollection{lecun93,
  author = 	 "Yann LeCun, Patrice Y. Simard, and Barak Perlmutter",
  title = 	 "Automatic Learning Rate Maximization by On-Line
                  Estimation of the Hessian's Eigenvectors",
  booktitle = 	 Tnips5,
  publisher = 	 "Morgan Kaufman",
  year = 	 1993,
  editor = 	 "Stephen Jose Hanson, Jack D. Cowan, and C. Lee Giles",
  pages = 	 "156-163",
  file=		 "6-9-93",

  annote =	 "Use the length of the first eigenvector of the
Hessian to determine the learning rate.  The local curvature of the
error space is greatest along the first eigenvector, to the extent
predicted by the eigenvalue.  The learning rate is limited by this.  

They propose a novel scheme for computing the this eigenvector.  The
goal is to compute the product of the Hessian with an arbitrary
vector.  Previous the product was approximated as the difference
between the gradient at the current weight vector and the gradient at
the weight vector plus a small component of the vector of interest.
While true for very small factors this process has several numerical
problems.

A number of other interesting application of such a technique are
proposed.  One could for example remap the weightspace along the
eigenvectors of the hessian.   This would make every error surface
more like a perfect bowl shape.

"
}

@Article{perlmutter93,
  author = 	 "Barak A. Perlmutter",
  title = 	 "Fast Exact Multiplication by the Hessian",
  journal = 	 "Neural Computation",
  year = 	 1993,
  file=		 "6-9-93",

  annote =	 "An extesion of the theory of the work described in
lecun93.  The goal is to compute the product of the Hessian with an
arbitrary vector.  Previous the product was approximated as the
difference between the gradient at the current weight vector and the
gradient at the weight vector plus a small component of the vector of
interest.  While true for very small factors this process has several
numerical problems.

Barak explicitly computes a novel differential operator R{} such that
R{\grad{w}} is the desired product.  He just completes the definition
of the derivative by taking the limit as the factor goes to zero.  He
applies this operator to several neural network schemes, with great
notational simplicity, and derives the computation of the Hessian
product very elegantly.

"
}

@article{MarNis78,
author = "David Marr and H.K. Nishihara",
title = "Representation and Recognition of the Spatial Organization of Three Dimensional Structure",
journal = prslb,
number = 200,
year = 1978,
pages = "269--294",
  file=		 "6-9-93",

  annote =	 "This is one of the papers that I read for my area exam.
It is an overview of their theories about matching objects to images."

}

@TechReport{CapGir90,
  author = 	 "Bruno Caprile and Federico Girosi",
  title = 	 "A Nondeterministic Minimization Algorithm",
  institution =  Ailab,
  year = 	 1990,
  number = 	 1254,

  file=		 "7-2-93",

  annote =	 "Finding parameters for an HBF can be a very
		  difficult process.  Gradient descent in error seems
		  to get caught on a variety of local minima.  They
		  suggest that one use a randomized best first search.
		  Keep the best set of parameters, make small random
		  changes, only accept better parameter sets.  If a
		  small change make the error larger, halve the size of
		  the random steps (or halve the variance).  If the
		  error gets smaller, double the size."
}


@Misc{MitHol93,
  author = 	 "Melanie Mitchell and John H. Holland",
  title = 	 "When Will a Genetic Algorithm Outperform Hill-Climbing",
  howpublished = neuroprose,
  year = 	 1993,
  month = 	 "June",
  file=		 "7-2-93",

  annote =	 "Embarrassingly, even for a fitness function that was
specifically designed to display the strengths of a GA, a randomized
best first search outperformed a GA.  This is an attempt to explain
this disappointment.  In the end of the paper the author's succeed in
redesigning the fitness function so that in some cases the GA is
better than random search.  

The randomized best first search that is used as a bench mark is
almost the same as Caprile and Girosi suggest \cite[CapGir90]."

}

@InCollection{AndSnoTreGra90,
  author = 	 "R. A. Andersen, R. J. Snowden, S. Treue, and M. Graziano",
  title = 	 "Hierarchical Processing of Motion in the Visual
		  Cortex of Monkey",
  booktitle = 	 "Cold Spring Harbor Symposia on Quantitative Biology",
  publisher = 	 "Cold Spring Harbor Press",
  year = 	 1990,
  pages =	 "741-748",

  file=		 "7-2-93",

  annote =	 "Interesting results in V1 versus MT.  Motion sensitive
cells in V1 ar not effected by the presence of additional moving
stimuli in their receptive fields -- as would be generated by
transparent motion.  V1 cells seems to select for a particular motion
and simply ignore other motions that are present.  MT cells are effect
by the presence of additional motion.  MT cells are depressed by the
presence of motion opposite preferred motion.

They also discuss the role of MST for transparent motion stimuli."

}

@InProceedings{OtsuKurita88,
  author = 	 "Nobuyuki Otsu and Takio Kurita",
  title = 	 "A New Scheme for Pratical Flexible and Intelligent
		  Vision Systems",
  booktitle = 	 "IAPR Workshop on Computer Vision",
  year = 	 1988,
  address = 	 "Tokyo",
  month = 	 "Oct",

  file=		 "8-11-93",

  annote =	 "This is
very interesting work and closely related to my own.  The results on
face recognition are good.  I am however very skeptical about the
particular features that they select.  They seem too simple to be
capable of a reasonable set of descriminations.

In their newer work the local features are non-linear, being the
product of pixels.  I can't imagine that this would generalize well
w.r.t to rotation and scale.  I use linear features, that should be
slightly more robust."

}

@article	( DESIMONEETAL84,
key	=	"Desimone et al." ,
author	=	"Robert Desimone and T. D. Albright and Charles G. Gross and 
		 C. Bruce" ,
title	=	"Stimulus selective properties of inferior temporal neurons 
		 in the macaque" ,
journal	=	JNeuSci ,
volume	=	"4" ,
year	=	"1984" ,
pages	=	"2051-2062" ,
keywords=	"neurophysiology, IT, face perception" ,
annote	=	"Systematic investigation of response properties of 
		 inferotemporal neurons." ,
summary	=	"110 (out of 151 total) visually responsive cell in IT were 
		 studied. 41% were unselective. Of the selective ones, 36% 
		 were selective for shape, 15% for color, and 42% were 
		 unclassified. Responses of 5 of the 151 cells were specific 
		 to particular objects (2 to hands, 3 to faces). More 
		 face-sensitive cells were found in the depths of the 
		 superior temporal sulcus. Thes object-specific cells were 
		 insensitive to changes in size, orientation (except for 
		 rotation in depth), and (relatively) in position. Removing 
		 components diminished (but did not eliminate) the responses. 
		 Scrambling severely reduced the response. Some 
		 face-sensitive cells preferred frontal views while others 
		 preferred profiles." ,
comments=	"Electrophysiological study of inferotemporal neurons (from 
		 Desimone et al., 1984)" 
)

@TechReport{WeiEde93,
  author = 	 "Yair Weiss and Shimon Edelman",
  title = 	 "Representation with receptive fields: gearing up for
		  recognition",
  institution =  "Weizmann Institute",
  year = 	 1993,
  number = 	 "CS-TR 93-09",
  month = 	 "August",
  file =         "8-30-93",

  annote =	 "Simplistic theoretical analysis of the utility of a
V1-like representation for object recognition.  They take a simple
model of whole-image template matching then show that a V1-like
representation should be better for generalization to novel views of
the object.  They do a few unconvincing experiments with unconvincing
results.  Still it is an interesting approach and related to mine.

As an example, with two template images the ``retinal'' representation
peforms better than the ``cortical'' representation.  This contradicts
their main claim."

}

@TechReport{Edelman93,
  author = 	 "Shimon Edelman",
  title = 	 "Representation, Similarity, and the Chorus of Prototypes",
  institution =  "Weizmann Institute",
  year = 	 1993,
  number = 	 "CS-TR 93-??",
  month = 	 "July",
  file =         "7-14-93",

  annote =	 "Argues that recognition and hence probably area IT is
done with a distributed representation.  There are certain persistance
representations say of object with which you are very familiar.  Other
objects, ephemeral objects, are represented based on the similarity to
these persistant objects.  To do this, one must build a recognition
system that has broad tuning curves for its object recognizers.  A
broad tuning curve assures that an ephemeral object makes several
persistant object recognizers respond significantly.  This joint
response vector is the representation for the ephemeral object.
Without broad tuning it would be likely that many ephemeral objects
would evoke no significant response from any persistant recognizer.
This would prevent recognition based on these responses.  

Interestingly a broad tuning curve makes recognition more difficult.
Broad tuning curves implies that the response of a detector for one
object to an object of another type will be non-zero.  This may make
the decision of which object is present more difficult."

}

@TECHREPORT{poggio89,
	AUTHOR = "Tomaso Poggio and Federico Girosi",
	TITLE = "A Theory of Networks for Approximation and Learning",
	INSTITUTION = "MIT",
	YEAR = 1989,
	TYPE = "AI Memo",
	NUMBER = 1140,
  file =         "9-15-93",

  annote =	 "They prove that GRBF have spline-like properties.  In
addition they prove that if there is a regularizer over the space of
approximating funtions (in essence proscribing that smooth functions
are more likely) then a GRBF network finds the most likely
approximation to the data.  This is very nice because at no point is
search of minimization necessary to find this function."

 }

@Misc{Atkeson91,
  author = 	 "Chris Atkeson",
  title = 	 "Memory-Based Approaches to Approximating Continuous Functions",
  howpublished = "Presented at a workshop and the Sante Fe Institute",
  year = 	 1991,
  month = 	 "March",
  file =         "9-15-93",

  annote = "This is one of Chris's general overviews of memory-based
techniques and his work on using memory based function approximation
of control.  Contains a huge number of references to the field."

}

@Misc{Brieman92,
  author =	 "Leo Breiman", 
  title =	 "Hinging Hyperplanes for Regression, Classification, and Function
Approximation",
  howpublished = "??",
  year =	 1992,
  file =         "9-15-93",

  annote =	 "Interesting and efficient technique for function
approximation.  Bears strong resemblances to Cascade Correlation of
Fahlman."

  }

@Article{WhiBal91,
  author = 	 "Steven D. Whitehead and Dana H. Ballard",
  title = 	 "Learning to Perceive and Act by Trial and Error",
  journal = 	 "Machine Learning",
  year = 	 1991,
  number =	 7,
  pages =	 "45-83",
  file =         "9-15-93",

  annote =	 "I haven't read it.  Contains an overview of the work done
for Steve's thesis where he uses reinforcement learning to learn to
percieve and act in a simple environment.  Perception is important
(where typically it is ignored) because the agent cannot sense the
entire state of the world just the state of a small part.  They make
many parallels to the work of Agre and Chapman -- arguably it is the
first real attempt to learn in an environment that is similar to the
ones that they use.  "

}

@TechReport{MosUll91,
  author = 	 "???",
  title = 	 "Limitations of Non Model-Based Recognition Schemes",
  institution =  Ailab,
  year = 	 1991,
  TYPE =         "AI Memo",
  number = 	 1301,

  file =         "9-15-93b",

  annote = "Interesting but somewhat abstract proof that there can be
no function of an image whose value uniquely identifies an object that
is invariant to object pose.  They are careful to state (but it
remains obscure) that there proof applies to functions that work for
all possible objects (i.e. there is no set of model objects).  

This proof does not hold, for instance, when the recognition function
can learn from a set of object instances.  Even though this function
may be able to learn to recognize any object this way.

The authors also make some strong simplifying assumptions about the
nature of an object and its image.

In my mind this proof is merely a curiosity.  Though there are people
who have attempted to build non-model based recognition schemes."

}

@Misc{viola92,
  author = 	 "Paul A Viola",
  title = 	 "A Review of Model Based Object Recognition",
  howpublished = "Area Exam at the MIT AI Lab",
  year = 	 1992,
  month = 	 "December",
  file =         "9-15-93b",

  annote = "Well there it is.  Sort of interesting overview.  But it
needs a lot of work.  If I need to write such a thing again this would
be a reasonable starting point."

}

@TechReport{EdeWei89,
  author = 	 "Shimon Edelman and Daphna Weinshall",
  title = 	 "A self-organizing multiple-view representation of 3D
		  objects",
  institution =  Ailab,
  year = 	 1989,
  TYPE =         "AI Memo",
  number = 	 1146,

  file =         "9-15-93b",

  annote = "The short answer is temporal coherency and a large memory
can build one a system that does object recognition (if you ignore
most of the real vision problem).

See note attached to paper."

  }

@TechReport{Fritzke93,
  author = 	 "Bernd Fritzke",
  title = 	 "Growing Cell Structures -- A Self-organized Network
		  for Unsupervised and Supervised Learning",
  institution =  icsi,
  year = 	 1993,
  number =	 "TR-93-026",
  month =	 "May",

  file =         "9-27-93",

  annote =	 "Interesting way of adding units to an RBF network.  One
must know the dimensionality of the data beforehand (i.e. that it is
2D data in the 10D space).  No plausible experiments in higher
dimensional spaces.

For supervised learning adds units where the error is high (sort of).
Uses something like Moody/Darken for RBF widths and weights.
"

}

@Misc{NowSej93,
  author =	 "Steven J. Nowlan and Terrence J. Sejnowski",
  title =	 "A Selection Model for Motion Processing in Area MT
		  of Primates",
  howpublished = "Submitted J.  Neuroscience",
  year =	 1993,
  month =	 "September",

  file =         "9-27-93",

  annote =	 "I read this for them.  I am still unclear on the
		  nature of the training set and the model's
		  dependance on this."

}

@Article{HutKlaRuc93,
  author = 	 "Daniel P. Huttenlocher, Gregory A. Klanderman and
		  William J. Rucklidge",
  title = 	 "Comparing Images Using The Hausdorff Distance",
  journal = 	 PAMI,
  year = 	 1993,
  volume =	 15,
  number =	 9,
  pages =	 "850-863",
  month =	 "September",

  file =         "9-27-93",

  annote =	 "Suggest a new distance metric for sparce binary images
(edge images).  Takes explicitly into account that a small
perturbation in the position of an on pixel should not penalize the
metric too much (like correlation does).  I'm not sure this is much
better that something like Ken Nakyama's sign of the DOG images.
Instead of just extracting the edges he keeps track of the sign of the
DOG of an image.  This is a dense image.  A correlation using these
images does not have the same problems.  Small shifts do not incur as
serious a penalty."

}

@TechReport{HutRuc93,
  author = 	 "Daniel P. Huttenlocher and William J. Rucklidge",
  title = 	 "A Multi-Resolution Technique for Comparing Images
		  Using the Hausdorff Distance.",
  institution =  "Department of Computer Science Cornell University",
  year = 	 1993,
  number = 	 "CUCS TR 92-1321",

  file =         "9-27-93",

  annote = "TR version of the PAMI paper.  Includes a technique that
deals with scale."

}


@Misc{Wong93,
  author = 	 "Yiu-fai Wong",
  title = 	 "A Nonlinear Scale-Space Filter by Physical Computation",
  howpublished = neuroprose,
  year = 	 1993,
  month = 	 "October",

  file =         "10-6-93",

  annote = "An alternative to aniosotropic difusion and line
processes.  Treat the image preprocessing as a clustering technique,
where the value of a pixel is the value of a robust cluster of pixels
about that point.  The process is a iterative one.  I believe (though
I'm not positive) that it may require significant computation to
compute this function."

 }

@Misc{Thau92,
  author =	 "Robert S. Thau",
  title =	 "Illuminant Percompensation for Texture
		  Discrimination using Filters",
  howpublished = "MIT BCS research project.",
  year =	 1992,
  month =	 "May",

  file =         "10-6-93",

  annote =	 "Interesting analysis of Malik and Perona's work on
texture descrimination.  Points out that much of the complexity of the
model is unecessary.  Also that the model does not work if there is
significant variation in lighting.  That for instance a system like
Malik's is more sensitive to changes in lighting than changes in
texture.  Thau suggests an illumination precompensation step that
seems to rescue the technique." 

 }

@Misc{NeaHin93,
  author = 	 "Radford M. Neal and Geofffrey E. Hinton",
  title = 	 "A New View of the EM Algorithm that Justifies
		  Incremental and Other Variants",
  howpublished = neuroprose,
  year = 	 1993,
  month = 	 "February",

  file =         "10-6-93",

  annote = "Didn't quite understand all of this.  Essentially creates
a single objective function that both E and M are gaurenteed to
reduce.  This way E and M can get mixed together."

}


@InCollection{SimLeCDen93,
  author = 	 "Patrice Y. Simard, Yann LeCun and John Denker",
  title = 	 "Efficient Pattern Recognition Using a New
		  Transformation Distance",
  booktitle = 	 Tnips5,
  publisher = 	 "Morgan Kaufman",
  year = 	 1993,
  editor = 	 "Stephen Jose Hanson, Jack D. Cowan, and C. Lee Giles",
  file=       "10-6-93",
 
  annote= "With a large dataset uses nearest neighbor to classify images
of handprinted digits.  They define a new distance metric that takes
apriori information about the expected deformations of the object into
account.  Essentially they define a set of vectors along which the
distance to a particular example should count for zero.  These vectors
are the derivatives of the example w.r.t.  things like scaling,
rotation and thickening.

The implementation first uses distance in pixel space to find a bunch
of close example.  Then the nearest example using the new metric is
the classification."


}

@TechReport{TorVio91,
  author = 	 "Mark Torrance and Paul Viola",
  title = 	 "The AGENT0 Manual",
  institution =  "Stanford University Department of Computer Science", 
  year = 	 1991,
  number = 	 "STAN-CS-91-1389",
  month = 	 "April",

  file=       "10-6-93",

  annote = 	 "Ack."
}

@Article{StoAlb93,
  author = 	 "Gene R. Stoner and Thomas D. Albright",
  title = 	 "Image Segmentation Cues in Motion Processing:
		  Implications for Modularity in Vision",
  journal = 	 jcogneuro,
  year = 	 1993,
  volume = 	 "5:2",
  pages = 	 "129-149",

  file=       "10-18-93",

  annote = "Describes a serious of very nice experiments where it is
demonstrated that motion perception of transparent stimuli is
dependant on physical interpretations of a scene.  i.e. the same
stimuli might appear as two transparent objects or a single object.  A
transparent interpretation stems from cues that imply a particular
physical interpretation.  These cues can be other than motion cues:
like stereo or intensity.  This contradicts the theory that vision
processing is primarily a separation of different fucntions like
stereo, shape from shading, and motion (ala Marr).

Cells in MT have been found whose activity is dependant on the
physical interpretation of the scene.  If the stimuli is possibly
transparent then the cell is a component motion detector.  If it is
not transparent the cell sees the pattern motion.
"

}

@Article{KimWil93,
  author = 	 "Jeounghoon Kim and Hugh R. Wilson",
  title = 	 "Dependence of Plaid Motion Coherence on Component
		  Grating Directions",
  journal = 	 visres,
  year = 	 1993,
  volume = 	 33,
  number = 	 17,
  pages = 	 "2479-2489",

  file=       "10-18-93",

  annote = "Show that component versus pattern motion percepts of
overlayed grating is very sensitive to the angle of relative motion.
Motions that a almost more colinear are more likely to cohere into a
pattern that motions that are opposite.  This seems to the dominant
phenomena when the motions are away from perpendicular.  

Also contains a simple theory that purports to explain the transparent
stimuli of Stoner and Albright.  This theory is directly refuted by
experiments described in \cite{StoAlb93}.  

"
}

@Proceedings{aaaifall93,
  title = 	 aaaisymp93,
  year = 	 1993,

  file=       "10-18-93",
 
  annote = "Should be superceded by the AAAI TR.  Don't know the number."

}

@Article{TreSat90,
  author = 	 "Anne Treisman and Sharon Sato",
  title = 	 "Conjunction Search Revisited",
  journal = 	 jexppsych,
  year = 	 1990,
  volume =	 16,
  number =	 3,
  pages =	 "459-478",

  file=       "10-18-93",

  OPTannote = "A reevaluation of the work on visual search and pop-out.
Treisman's original theory that objects with fundamental difference
might pop-out and that objects defined by the conjunctions of these
differences has been called in to serious question.  Essentially other
researcher have not found a cut and dried distinction -- most
especially they have not found a 2 to 1 slope ratio between target
present and target absent presentations.

Through a series of experiments Treisman and Sato look to quantify
what is going on.  Like much of the data the paper is difficult to
interpret.  Carefully describing experiments without clear
conclusions.

In relation to the work of Stoner and Albright, pop-out may be a
side-effect of a system that interprets images in which two objects
can be present with their spatial parts interspersed (say a lion
moving behind a tree).  In this case the original pop-out experiments
can be interpreted as grouping the distractors into evidence (more
than likely textural) of a single object.  In that case it is easy to
see if the target is present -- since there are only two objects the
target and the distractors .  The stimuli that trigger serial search
are made up of tokens that cannot be grouped easily into two different
objects (probably because the textural grouping system is limited).
Treisman finds parallel evidence that the ease of search is related to
the easy of texture segmentation.  Also she demonstrates that when the
distractor take on a more complex and varied form (including more than
2 different features) search times increase.  This is predictable from
the textural a segmentation theory.  Increasing the number of types of
distractors increases the number of different objects that must be
``seen''.
"

}

@InBook{Nakayama90,
  author = 	 "K. Nakayama",
  title = 	 "Vision Coding and Efficiency",
  chapter = 	 "The iconic bottleneck and the tenuous link between
		  early visual processing and perception",
  publisher = 	 "Cambridge University Press",
  year = 	 1990,
  editor =	 "Colin Blakemore",
  pages =	 "411-422",

  file=       "10-18-93",

  OPTannote = "This theory of object recognition is driven by the
assumption that representations of visual objects do not contain a lot
of information.  He is unclear about his argument for this.  It may be
because objects need to be simple (memory requirements) or that there
is a limited amount of bandwidth between memory and visual processing
(the iconic bottleneck).  He argues for scale invariance by claiming
there is a pyramid representation in the brain (which there may be but
it is clear that different types of information are represented at
different levels).  The same memory can match at different levels of
the pyramid yielding a kind of scale invariance.

Because of the very low bandwidth of the bottleneck Nakayama argues
that recognition processing is a serial process.  It is not clear that
this is supported by psychophysics and neuroscience.  It seems that
people can recognize objects with very short presentations.  Similarly
cells don't seems to need to be serialized.  They appear to work in
parallel thoughout the visual field.  \cite{chelazzi93} seems to argue
that the only time cells get in each other's way is when the share a
receptive field.  Here the interaction seems inhibitory.

The general form of his arguments define a system that would be a
perfect application of some sort of reinforcement learning (using say
TD-lambda).

" }

@Article{PosPet90,
  author = 	 "Michael I. Posner and Steven E. Pertersen",
  title = 	 "The Attention System of the Human Brain",
  journal = 	 annrevneuro,
  year = 	 1990,
  volume =	 13,
  pages =	 "25-42",

  file=       "10-18-93",

  annote =	 "List of attentional mechanisms in the human brain.  Not
much interesting discussion."

}


@Misc{MatKon93,
  author = 	 "T. Matsumoto and K. Kondo",
  title = 	 "Realisation of ``Weak Rod'' by a Double Layer
		  Parallel Network",
  howpublished = "Reviewers copy",
  year = 	 1993,

  file=       "11-15-93",

  annote = 	 "Included is the second version of the paper the
		  reviewers comments on the first version.  Careful
		  don't ever reference this."
}



@Article{Motter93,
  author = 	 "Brad C. Motter",
  title = 	 "Focla Attention Produces Spatially Selective
		  Processing in Visual Cortical Areas V1, V2 and V4 in
		  the Presence of Competing Stimuli",
  journal = 	 jneusci,
  year = 	 1993,
  volume =	 70,
  number =	 3,
  pages =	 "909-919",
  month =	 "September",
  file=       "11-15-93",

  annote =	 "Role of attention on visual cortex cells.  Attention
seems to have some effect even if it is outside the receptive field of
a cell.  But, the effect is sometimes excititory and sometimes
inhibitory.  This is one of few papers to show any effect in V1 from
attention.  "

}

@Misc{LeoMii93,
  author = 	 "W. K. Leow and R. Miikkulainen",
  title = 	 "Representing and Learning Visual Schemas in Neural
	Networks for Scene Analysis",
  howpublished = "Reviewers copy",
  year = 	 1993,

  file=       "11-15-93",

  annote = "I reviewed this for NC.  See
		  reviews/nc-leow-miikkulainen.tex.  Careful don't
		  ever reference this."
}

@Article{Tanaka,
  author = 	 "Keiji Tanaka",
  title = 	 "Laboratory for Neural Information Processing",
  journal = 	 "??",
  year = 	 "??",

  file=          "1-3-94",

  annote =	 "Great compendium of experiments in IT.  Lots of
		  data.  This is a pre-print.  I don't know where it
		  ended up."
}


@InProceedings{Nishihara89,
  author = 	 "H. K. Nishihara",
  title = 	 "Psychophysical and Computational Test Comparing the
		  Sign-Correlation and Zero-Crossing Models of Human
		  Stereo Vision",
  booktitle = 	 "Image Understanding and Machine Vision: 1989
		  Technical Digest Series",
  year = 	 1989,
  pages =	 "40-43",
  organization = "Optical Society of American",

  file = 	 "1-3-94",

  annote = " Using his sign correlation theory of vision he shows that a
a transparent random dot stereo pair can be correctly precieved.  He
also shows that a purely zero-crossing theory could not correctly
perceive the image."


}


@TechReport{Nishihara91,
  author = 	 "H. K. Nishihara",
  title = 	 "Minimal Meaningful Measurement Tools",
  institution =  "Teleos Research",
  year = 	 1991,
  number =	 "TR-91-01",
  address =	 "Teleos Research, 576 Middlefield Road, Palo Alto, CA
		  94301",

  file = "1-3-93",

  annote =	 "Position paper on the nature of useful low-level vision
tools.  Analyzes tools as if they were intended to be used by a blind
person.  Quick overview of the prism stereo system that uses his sign
of the DOG algorithm.  "

 }


@Article{Nishihara87,
  author = 	 "H. K. Nishihara",
  title = 	 "Practical Real-Time Imaging Stereo Matcher",
  journal = 	 "Optical Engineering",
  year = 	 1987,
  volume = 	 23,
  number = 	 5,
  pages = 	 "536-545",
  month = 	 "September/October",

  file = "1-3-93",

  annote = "Overview of practical stereo to date, 1987, and a
description of his algorithm that uses the sign of the DOG of a stereo
pair to find disparity.  He argues that the sign of DOG is less
sensitive to noise than the zero-crossing.  For instance the total
numbers of bits that will be flipped is low for even moderately large
image noise.  Conversely, the zero-crossing will more a significant
distance under the influence of noise.

Interesting properties: i) The autocorrelation of the sign of DOG is
sharp, the autocorrelation of the DOG is not.  ii) The autocorrelation
of the sign of DOG can be made less sharp, providing for multi-scale
search.  iii) Sign of DOG is variant under lighting changes: intensity
and contrast.  DOG doesn't give you contrast.

There is one hole in this paper: a real analysis of DOG zero-crossing
matching is missing.  It is clear that the zero-crossings move but not
by how much.

Finally the resulting matching computation, binary correlation, is
easily implemented in hardware."

}


@InProceedings{MurNay93,
  author = 	 "Hiroshi Murase and Shree K. Nayar",
  title = 	 "Learning and Recognition of 3-D Objects from Brightness Images",
  booktitle = 	 aaaisymp93,
  year = 	 1993,
  organization = "AAAI",

  file=       "10-18-93",

  annote = 	 "Neat approach, but it is unclear if principle
		  components are the right thing.  Fundamentally
		  assumes segmentation."
}

@InCollection{Ahmad94,
  author = 	 "Subutai Ahmad",
  title = 	 "Feature Densities are Required for Computing Feature
                  Correspondances",
  booktitle = 	 Tnips6,
  publisher = 	 "Morgan Kaufman",
  year = 	 1994,
  editor = 	 "J D. Cowan, G. Tesauro and J. Alspector",
  file=       "2-14-94",
 
  annote= "Interesting in that it touches on some things I have been
thinking about.  Essentially you need to integrate over all possible
correspondances to pick the correct object in an image.  Typically one
only picks the best correspondance and evaluates it."  

}

@MastersThesis{viola90,
  author = 	 "Paul A. Viola",
  title = 	 "Adaptive Gaze Control",
  school = 	 "MIT",
  year = 	 1990,
  month = 	 "October"
}



@Article{Desimone90,
  author = 	 "R. Desimone, M. Wessinger, L. Thomas and W. Schneider",
  title = 	 "Attentional Control of Visual Perception: Cortical
		  and Subcortical Mechanisms",
  journal = 	 "Cold Spring Harbor Symposia on Quantitative Biology",
  year = 	 1990,
  volume = 	 "LV",
  pages = 	 "963-971",
  file=       "2-14-94",
 
  annote= "Pulvinar and Superior colliculus play a role in attention.
    Suppressing these area seem to impair the ability to do descrimination
    in the presence of distractors.  The tasks are spatially cued: the
    animal is asked to do a match to sample at a particular location.
    Distractors appear at other locations.  No deficit is seen in the
    absence of distractors.  These areas seem to help a particular retinal
    location win a competition to be represented at a higher level.
    Without them the distractors win.  By inhibiting local regions of S.
    colliculus particular area of the retina can have attention inhibited."
 
 }


@InCollection{GolMjoRan94,
  author = 	 "Steven Gold, Eric Mjolsness and Anand Rangarajan",
  title = 	 "Clustering with a Domain-Specific Distance Measure",
  booktitle = 	 Tnips6,
  publisher = 	 "Morgan Kaufman",
  year = 	 1994,
  editor = 	 "J D. Cowan, G. Tesauro and J. Alspector",
  file=       "2-14-94",
 
  annote= "See note on the cover of the paper.  Basically this is an
obvious redefinition of the vision problem with the suggestion that
the right way to solve the correspondance/pose problem is with
gradient descent.  This is guarenteed not to work."

}

@InCollection{ChiMjo94,
  author = 	 "Chien-Ping Lu and Eric Mjolsness",
  title = 	 "Two-Dimensional Object Localization by
		  Coarse-to-Fine Correlation Matching",
  booktitle = 	 Tnips6,
  publisher = 	 "Morgan Kaufman",
  year = 	 1994,
  editor = 	 "J D. Cowan, G. Tesauro and J. Alspector",
  file=       "2-14-94",
 
  annote= "Just like \cite{GolMjoRan94} this is a hunk of garbage.  Once
again try to solve the corr/pose problem with gradient descent.
Because this is really hard w.r.t. rotation they simply make a lot of
copies of the model at the different rotations.  Luckily for him their
is only one rotation parameter."

}

@TechReport{Beymer93,
  author = 	 "David Beymer",
  title = 	 "Face Recognition Under Varying Pose",
  institution =  mitai,
  year = 	 1993,
  TYPE = "AI Memo",
  number =	 1461,
  month =	 "December",
  file=       "2-14-94",

  annote =	 "Good description of Davids work.  Two stage process
estimate pose using general templates and then estimate identity by
comparing faces with faces in that pose.  Pose step is expensize and
not very effective.  Uses optical flow to bring the faces into closer
alignment."  

}

@Article{Wehner87,
  author = 	 "Rudiger Wehner",
  title = 	 "``Matched filters'' --- neural models of the
		  external world",
  journal = 	 "Journal of Comparative Physiology A",
  year = 	 1987,
  number =	 161,
  pages =	 "511-537",
  file=          "2-14-94",

  annote =	 "Very neat paper about how many types of animals use
simple filters to compute what they need to know.  Like the distance
to prey or the position of preditors.  "

 }


@InCollection{Wells65,
  author = 	 "M. J. Wells",
  title = 	 "What the Octopus Makes of It: Our World from Another
		  Point of View",
  booktitle = 	 "Readings in Animal Behavior",
  publisher = 	 "Holt, Rinehart and Winston",
  year = 	 1965,
  editor =	 "Thomas E. McGill",
  chapter =	 31,
  pages =	 "338-346",
  file=          "3-21-94",

  annote =	 "Describes learning in the octopus and its dependance on
sensory processing.  Includes visual and tactile learning." 

 }

@InCollection{Land72,
  author = 	 "M. F. Land",
  title = 	 "Mechanisms of Orientation and Pattern Recognition by
		  Jumping Spiders (Salticidae)",
  booktitle = 	 "Information Processing in the Visual Systems of Arthropods",
  publisher = 	 "Springer-Verlag",
  year = 	 1972,
  editor =	 "R. Wehner",
  chapter =	 8,
  file=          "3-21-94",

  annote =	 "Neat description of the simplicity of the spider visual
system.  What is surprising is that it is quite functional."

  }


@Article{Desimone91,
  author = 	 "Robert Desimone",
  title = 	 "Face-Selective Cells in the Temporal Cortex of Monkeys",
  journal = 	 jcogneuro,
  year = 	 1991,
  volume =	 3,
  number =	 1,
  pages =	 "1-8",
  file=          "3-21-94",

  annote =	 "Brief overview of the literature on face recogntion cells
in IT and STP.  Useful in that it descibes what is known about the
architecutre of the superior temporal sulcus which is the border
between STP and IT.  Presents two hypotheses: that faces are special
to the brain, that faces are simply more important and common, that
may describe why such a large number of object sensitive cells seem to
code exclusively for faces."

}





@Misc{DayZem94,
  author =	 "Peter Dayan and Richark S. Zemel",
  title =	 "Competition and Multiple Cause Models",
  howpublished = "Review Copy",
  year =	 1994,
  file=          "3-21-94",

  annote =	 "Nice rehashing of Eric Saunds work on Noisy-OR.  By
		  using the activation function of Keeler, Rumelhart
		  and Leow along with a reconstruction evaluation
		  metric they get a representation that is sparse yet
		  not to sparse."
}



@Article{FreAde91,
  author = 	 "William T. Freeman and Edward H. Adelson",
  title = 	 "The Design and Use of Steerable Filters",
  journal =	 PAMI,
  year =	 1991,
  volume =	 13,
  number =	 9,
  pages =	 "891-906",
  month =	 "September",
  file=          "4-29-94",
 
  annote= "The reference for steerable filters and oriented energy.
Also used there stuff to do edge detection."

}

@Article{BurAde83,
  author = 	 "P.J. Burt and E. H. Adelson",
  title = 	 "The Laplacian pyramid as a compact image code",
  journal =	 ieee-com,
  year =	 1983,
  volume =	 31,
  number =	 4,
  pages =	 "532-540",
  file=          "4-29-94",
 
  annote= "I don't have this."
}

@Article{BruPog93,
  author = 	 "Roberto Brunelli and Tomaso Poggio",
  title = 	 "Face Recognition: Feature versus Templates",
  journal =	 PAMI,
  year =	 1992,
  volume =	 15,
  number =	 10,
  pages =	 "1042-1052",
  month =	 "October",
  file=          "4-29-94",
 
  annote= "See my notebook.  Interesting and very pragmatic.  Not really
		  ground-breaking however."

}

@Misc{EglDriRaf93,
  author =	 "Robert Egly and Jon Driver and Robert D. Rafal",
  title =	 "Shifting Visual Attention between Objects and
		  Locations: Evidence from Normal and Parietal Lesion Subjects",
  howpublished = "Pre-print",
  year =	 1993,
  file=          "5-11-94",
 
  annote= "Parietal seems to play a role in shifting attention from
location to location and from object to object.  But the question of
whether these are the same things is still open.  This paper attempts
to differentiate them.  Using lesion patients they reproduce that
parietal lesions slow shifts of attention between locations.  But they
also show that while right-hemi patients are equally impared in the
location vs. object tests, the left-hemisphere patients are much at
shifting between objects than locations." 
 }

@Misc{EglDriSta93,
  author =	 "Robert Egly and Jon Driver and Yves Starrveveld",
  title =	 "Covert Orienting in the Split-brain Reveals
		  Hemispheric Specialization for Object-based Attention",
  howpublished = "Pre-print",
  year =	 1993,
  file=          "5-11-94",
 
  annote= "Similar experiemnt to \cite{EglDriRaf93}.  But using a
split brained patient.  Using the usual single hemi presentations they
show that the left hemi is slower to shift attention between different
objects than different locations on the same object."

 }

@book{CoverThomas,
 YEAR = 1991,
 TITLE = "Elements of Information Theory",
 AUTHOR = "T.M. Cover and J.A. Thomas",
 PUBLISHER = "John Wiley \& Sons" ,
 ANNOTE = "Text book on everything about information theory." 
}

@Article{Adelson93,
  author = 	 "Edward H. Adelson",
  title = 	 "Perceptual Organization and the Judgement of Brightness",
  journal =	 "Science",
  year =	 1993,
  volume =	 262,
  pages =	 "2042-2044",
  month =	 "December",
  file=          "5-11-94",
 
  annote= "The best demonstration of perceptual, or perceived,
brightness being very different from actual brightness.  In this case
is shows that perception is guided by the 3D interpretation of the
scene.  No local processing, like retinex, could explain the
differences."

}

@article{NC:Foldiak91,
  author = {F\"oldi\'ak, P.},
   title = {Learning Invariance from Transformation Sequences},
    type = {Letter},
 journal = {Neural Computation},
  volume = {3},
  number = {2},
   pages = {194--200},
    year = {1991}
}



@TechReport{WanAde93,
  author = 	 "John Y. A. Wang and Edward H. Adelson",
  title = 	 "Layered Representation for Motion Analysis",
  institution =  media-vision,
  year = 	 1993,
  number =	 221,
  month =	 "April",

  file =         "5-11-94",

  annote =	 "This is the edge versus region debate all over again.
Most motion estimation algorithms attempt to smooth their noisy
estimates, except across edges or discontinuities.  Authors believe
that it is best to divide the image into regions with a single 2D
affine velocity.  They have a greedy iterative algorithm that assigns
pixels to these regions based on there fit."

}

@TechReport{DarSim93,
  author = 	 "Trevor Darrell and Eero Simoncelli",
  title = 	 "Separation of Transparent Motion into Layers using
		  Velocity-Tuned Mechanisms",
  institution =  media-vision,
  year = 	 "1993",
  OPTnumber = 	 "244",
  OPTmonth = 	 "October",

  file =         "5-11-94",

  OPTannote = "The outputs of a group of simple sine cosine oriented
energy detector are alway unimodal.  Higher order detectors, higher
derivatives of Gassians, are not uni-modal.  Using a third-derivative
steerable ensemble, they attempt to find transparent motion.  Where
there is transparancy there will be two distinct modes of the energy
distribution in orientation.  They use a scheme that is similar to
\cite{WanAde93} to segment to the images into regions of differing
motion.  There scheme is different because a single pixel can belong
to more than one layer/velocity."

}


@Misc{VecFar93,
  author =	 "Shaun P. Vecera and Martha J. Farah",
  title =	 "Attentional Selection from Spatially Invariant
		  Object Recognition",
  howpublished = "Submitted to Psychological Science",
  year =	 1993,
  file =         "5-11-94",

  annote = "Uses the Duncan experiment where two different objects are
displayed, and two properties of the ensemble is reported.  When the
properties come from the same object things go quicker--or are more
accurate--than when they come from different objects.  The authors
separated the objects spatially and showed that there was little
effect, same was still better than different.  In the end I wasn't
sure that the authors really proved anything."

}

@Article{Saund94,
  author = 	 "Eric Saund",
  title = 	 "A Multiple Cause Mixture Model for Unsupervised Learning",
  journal =	 "Submitted to Neural Computation",
  year =	 1994,
  file =         "5-21-94",

  annote = "When you look at the hidden unit of a network as being
binary variables that may cause an output unit to be on then the
combination function of the output unit is dependant on your
generative model.  For instance if you believe all the units must be
on, one of the units must be on, or many must be on.  A sigmoid
activation function has a fairly weird interpretation: if some
combination of units are on then the output should be on.  The error
function continues to push on the hidden units to make their
probabilities larger and larger.  Saund suggests that it may be
valuable to use a different model, where only one of units needs to be
on.  This allows the other hidden units to be ambiguous about the
input."

}

@Misc{LarSteVioSejHagEkm94,
  author =	 "Jan Larsen and Marian Steward Bartlett and Paul A.
		  Viola and Terrence J. Sejnowski and Joe Hager and
		  Paul Eckman",
  title =	 "Comparing Methods for Measuring Facial Expressions",
  howpublished = "Nips Submission",
  year =	 1994,
  file =         "5-21-94",

  annote = "foo"

}

@InProceedings{BuhLadMal90,
  author = 	 "Joachim Bumann and Martin Lades and Christoph von
		  der Malsburg",
  title = 	 "Size and Distortion Invariant Object Recognition by
		  Hierarchical Graph Matching",
  pages = "411--416",
  volume = "II",
  crossref = "NN:ijcnn89",
  file =         "5-21-94",

  annote = "Very similar to CFR.  For every model object they build a
grid at every point of which is a Gabor Jet.  To match a novel image,
they plop the grid of Jets down and move them around using a local
random search.  At each time they are maximizing a sum of the jet
match and the grid deformation.  Initially they only match the low res
Gabor functions.  As time progresses they match higher and higher
resolutions.  The losing part: its got to be very sensitive to initial
conditions, because there must be a load of local minima in the
objective function, but there is no mention of this.  Also they do not
attempt to learn anything about the Jets -- which are good, which are
bad -- and they are proud of it.  They believe this brings about
efficiency -- which it does.  

They compute the Jets everywhere and cache them, then moving a grid
point is cheap, they only need to compare a few numbers.  This might
not be too inefficient.

There is no discussion of the merits of Jet matching.  How expressive
are they?  How selective?"

}

@Misc{KalBieCoo94,
  OPTauthor = 	 "Peter Kalocsai and Irving Biederman and Eric E. Cooper",
  OPTtitle = 	 "To What Extent Can the Recognition of Unfamiliar
		  Faces be Accounted for by the Direct Output of
		  Simple Cells",
  OPThowpublished = "Poster at Association for Research in Vision and Ophthalmology",
  OPTyear = 	 "1994",
  OPTnote = 	 "Sarasota, Florida",
  file =         "5-21-94",

  annote = "They did some psychophysics on impoverished -- no hair --
face recognition under changes in pose and expression.  They obtained
error and reaction times.  They then ran the system in
\cite{BuhLadMal90} do obtain matching coefficients.  They correlations
between error/reaction times and matching coefficients was high in
some case -- .60 to .90.  

While correlation is high, this is a very contrived experiment.  I
believe these people were unknown to the subjects.  Presentation times
were very low -- 100 msecs.  With longer presentation times or with
faces that were known to the subjects, I am sure performance would have
been much better."

}

@Article{OlsAndVan93,
  author = 	 "Bruno A. Olshausen and Charles H. Anderson and David
		  C. Van Essen",
  title = 	 "A Neurobiological Model of Visual Attention and
		  Invariant Pattern Recognition Based on Dynamic
		  Routing of Information",
  journal =	 "Journal of Neuroscience",
  year =	 1993,
  volume =	 13,
  number =	 11,
  pages =	 "4700--4719",
  file =         "5-21-94",

  annote = "A big paper on Bruno's implementation of the shifter
hypothesis.  While the paper claims that the theory is neurobiological
it seems to be much more of an engineering paper." 

 }

@Article{NieKocRos93,
  author = 	 "Ernst Niebur and Christof Kock and Christopher Rosin",
  title = 	 "An Oscillation-Based Model of the Neuronal Basis of Attention",
  OPTjournal = 	 "Vision Research",
  OPTyear = 	 "1993",
  OPTvolume = 	 "33",
  OPTnumber = 	 "18",
  OPTpages = 	 "2789--2802",
  file =         "5-27-94",

  annote = "They claim that the firing of low level -- V1 etc -- cells
are tagged with by modulating their firing at 40 Hz.  The modulation
component is quite small.  They claim it may have been missed by
experimenters.  Some higher level cells are frequency tuned, and gate
subsequent processing so that it concentrates soley on the tagged
inputs.  --- Not a lot of evidence, some simulations, interesting
hypothesis." 

 }


@PhdThesis{Chaney93,
  author = 	 "Ronald Dean Chaney",
  title = 	 "Feature Extraction Without Edge Detection",
  school = 	 mitai,
  year = 	 "1993",
  file=          "On the shelf with other theses.",
 
  annote=        "Haven't really read."
}



@Article{Yuille91,
  author = 	 "Alan L. Yuille",
  title = 	 "Deformable Templates for Face Recognition",
  OPTjournal = 	 jcogneuro,
  OPTyear = 	 "1991",
  OPTvolume = 	 "3",
  OPTnumber = 	 "1",
  OPTpages = 	 "59--70",
  file=          "5-27-94",
 
  annote= "Presents the idea of using deformabel templates to fit
faces.  Deformable templates are really very much like generative
models that have a few -- 10 or so -- spatial parameters.  The edges
of the template like to generate edges near them, etc.  The templates
are fit using gradient descent---which seems problematic, though they
claim it isn't.  There are no overall experiements described however,
just a few templates and a few isolated examples.  How well does the
system work??"

}


@Article{MohNev92,
  author = 	 "Rakesh Mohan and Ramakant Nevatia",
  title = 	 "Perceptual Organization for Scene Segmentation and Description",
  journal =	 pami,
  year =	 1992,
  volume =	 14,
  number =	 6,
  pages =	 "616-635",
  file=          "5-27-94",


  OPTannote = 	 "For me this is a was puzzling paper.  They show excellent results on
bottom up segmentation of edge images of man made objects.  They use
no information besides the geometric relationships of the edges---no
contrast, texture or color is used.  This runs counter to all my
intuitions, that it is impossible to do this well.  To there credit
they point out that people can segment these edge images very well.
But it is not clear that it is easy to insure that people do not use
model information to assist them, given that people have a seemingly
infinite number of models available.

Essentially lines are grouped based on cocurvilinearity---they can be
linked to form a reasonably smooth curve. Then curve are paired to
form skeletons, that will form the eventual basis for the
identification of ribbons.  The skeletons can conflict.  Conflict
resolutions is done with a ``Hopfield'' network that relaxes to a
state where only the best skeletons are selected (longest, most
parallel with their bounding curves etc.).  Finally the curves and
their skeleton are capped to form ribbons.  The ribbons are then
grouped into objects on the basis of shared curves.

There are many parameters in this system, only a few of which have
been tested for sensitivity.  Only 3 experiments are shown, and only
20 have been done.  I wonder how reliable their results are?  But
there results are quite tantalizing."

}

@Book{Press92,
  author =       "William H. Press, Brian P. Flannery, Saul
		  A. Teukolsky and William T. Veterling", 
  title = 	 "Numerical Recipes in C: The Art of Scientific Computing",
  publisher = 	 "Cambridge University Press",
  year = 	 1992,
  address =	 "Cambridge, England",
  edition =	 "second edition"
}

@Article{JonPog95,
  author = 	 "Michael Jones and Tomaso Poggio",
  title = 	 "Model-based Matching of Line Drawings by Linear Combinations of
Prototypes",
  journal =	 iccv,
  year =	 1995,
  annote =       "Currently in my thesis stack of papers."

}


@Article{RobMun51,
  author = 	 "H. Robbins and S. Munroe",
  title = 	 "A stochastic approximation method",
  journal =	 "Annals of Mathematical Statistics",
  year =	 1951,
  volume =	 22,
  pages =	 "400-407",

  annote =	 "I don't have this but it is referenced in Ljung"
}

@Article{Shannon48,
  author = 	 "C. E. Shannon",
  title = 	 "A mathematical theory of communication",
  journal =	 "Bell Systems Technical Journal",
  year =	 1948,
  volume =	 27,
  pages =	 "379-423 and 623-656",

  annote =	 "I don't have this.  Got the reference from Haykin"
}

@article{Linsker88,
       author = "Linsker, R.",
        title = "Self-Organization in a Perceptual Network",
        pages = "105--117",
      journal =  computer,
        month = "March",
         year =  1988
  }

@MASTERSTHESIS{mellor95,
	AUTHOR = {J.P. Mellor},
	TITLE = {Enhanced Reality Visualization in a Surgical Environment},
	SCHOOL = {Massachusetts Institute of Technology},
	YEAR = 1995}

@INPROCEEDINGS{mellor95a,
	AUTHOR = {J.P. Mellor},
	TITLE = {Realtime Camera Calibration for Enhanced Reality Visualization},
	BOOKTITLE = {Computer Vision, Virtual Reality and Robotics in Medicine},
	PAGES = {471-475},
	MONTH = {April},
	YEAR = 1995,
	NOTE = {Nice, France}
}


@Article{WellsEtc96,
  author = 	 "William M. {Wells III} and Paul Viola and Hideki
		  Atsumi and Shin Nakajima and Ron Kikinis",
  title = 	 "Multi-Modal Volume Registration by Maximization of Mutual
Information",
  journal =	 "Medical Image Analysis",
  year =	 1996,
  volume =	 1,
  number =	 1
}


@Article{Tincher93,
  author = 	 "M. Tincher, C. R. Meyer, R. Gupta and D. M. Williams",
  title = 	 "Polynomial modeling and reduction
of RF body coil spatial inhomogeneity in MRI",
	journal = "IEEE Trans. Med. Imaging",
  year =	 1993,
  volume =	 12,
  number =	 2,
  pages =	 "361-365",
  annote =	 "From Sandy.  I don't have this paper."
}

@Unpublished{BRB,
  author = 	 "Kenneth Baclawski and Gian-Carlo Rota and Sara Billey",
  title = 	 "Introduction to the Theory of Probability",
  note = 	 "MIT course notes for 18.313",
  year =	 1990
}

@Article{Dempster77,
	AUTHOR = "A.P. Dempster and N.M. Laird and D.B. Rubin",
	TITLE = {{M}aximum {L}ikelihood from {I}ncomplete {D}ata via the {EM} {A}lgorithm},
  journal =	 "Journal of the Royal Statistical Society, Series B",
	YEAR = "1977",
	VOLUME = "39",
	PAGES = "1 -- 38",
	}



@article{NC:Jacobs91,
  author = {Jacobs, R.A. and Jordan, M.I. and Nowlan, S.J. and Hinton, G.E.},
   title = {Adaptive Mixtures of Local Experts},
    type = {Letter},
 journal = {Neural Computation},
  volume = {3},
  number = {1},
   pages = {79--87},
    year = {1991}
}

@inproceedings{VioWel95,
     author = "Paul A. Viola and William M. {Wells III}",
      title = "Alignment by Maximization of Mutual Information",
      pages = "16--23",
  booktitle = "Fifth Intl.\ Conf.\ on Computer Vision",
    address = "Cambridge, MA",
  publisher = "IEEE",
       year =  1995
}


@PhdThesis{ViolaPhD,
author = "Paul A. Viola",
title = "Alignment by Maximization of Mutual Information.",
school = "Massachusetts Institute of Technology",
year = "1995",
	note = "MIT AI Laboratory TR 1548"
}

@article{NN:Becker92,
     author = "Suzanna Becker and Geoffrey E. Hinton",
      title = "A Self-Organizing Neural Network
               That Discovers Surfaces in Random-dot Stereograms",
      pages = "161--163",
    journal = "Nature",
     volume =  355,
       year =  1992
}

@Article{Intrator92,
  author = 	 "Nathan Intrator and Leon N. Cooper",
  title = 	 "Objective Function Formulation of the BCM Theory of
		  Visual Cortical Plasticity: Statistical Connections,
		  Stability Conditions",
  journal =	 "Neural Networks",
  year =	 1992,
  volume =	 5,
  pages =	 "3-17"
}


@Article{Huber85,
  author = 	 "Peter J. Huber",
  title = 	 "Projection Pursuit",
  journal =	 "The Annals of Statistics",
  year =	 1985,
  volume =	 13,
  number =	 2,
  pages =	 "435--475",
  annote =	 "Look in the Projection Pursuit File"
}

@Article{Friedman87,
  author = 	 "Jerome H. Friedman",
  title = 	 "Exploratory Projection Pursuit",
  journal =	 "Journal of the American Statistical Association",
  year =	 1987,
  volume =	 82,
  number =	 397,
  pages =	 "249--266",
  annote =	 "Look in the Projection Pursuit File"
}



@Misc{WelVio95,
     author = "William M. {Wells III} and Paul A. Viola",
  title = 	 "Multi-Modal Volume Registration by Maximization of Mutual Information",
  howpublished = "In preparation.",
  year =	 1995
}



@inproceedings{malandain-al94a,
author="Malandain, G. and {Fern\'{a}ndez-Vidal}, S. and Rocchisani, J.M.",
title="Improving Registration of {3-D} Medical Images Using a Mechanical
        Based Method",
booktitle="3rd European Conference on Computer Vision (ECCV'94)",
address="Stockholm, Sweden",
year="1994",
month="May 2--6",
pages="131--136",
note="Lecture Notes in Computer Science 801",
rapact="1994"
}

@techreport{malandain-al95a,
author="Malandain, G. and {Fern\'{a}ndez-Vidal}, S. and Rocchisani, J.M.",
title="Physically Based Rigid Registration of {3-D} Free-Form Objects: Application to Medical Imaging",
institution="INRIA",
address="2004 route des Lucioles BP 93, 06902 Sophia Antipolis Cedex, France",
number="2453",
month="January",
year="1995",
note="submitted to CVGIP-IU"
}

@article        ( SWAINBALLARD91,
key	=	"Swain" ,
author	=	"Michael J. Swain and Dana H. Ballard" ,
title	=	"Color indexing" ,
journal	=	ijcv ,
volume	=	"7" ,
pages	=	"11-32" ,
year	=	"1991" ,
keywords=	"computer vision" 
)

@Article{ScaPen95,
  author = 	 "Stan Sclaroff and Alex P. Pentland",
  title = 	 "Modal Matching for Correspondence and Recognition",
  journal =	 pami,
  year =	 1995,
  volume =	 17,
  number =	 6,
  pages =	 "545--561"
}


@inproceedings{RaoBal95,
  author = 	 "Rajesh P. N. Rao and Dana H. Ballard",
  title = 	 "Object Indexing using an Iconic Sparse Distributed Memory",
  booktitle = iccv,
  publisher = "IEEE, Washington, DC",
  address = "Cambridge, MA",
  month = "June",
  year = 	 1995,
  pages =	 "24--31"
}


@inproceedings{Wells91,
	author = "William M. {Wells III}",
	title = "{MAP} {M}odel {M}atching",
	booktitle = "Proceedings of the Computer Society Conference on Computer Vision and Pattern Recognition",
	year = "1991",
	month = "June",
	pages = "486--492",
	organization = "IEEE",
	address = "Lahaina, Maui, Hawaii"
	}

@inproceedings{MRCAS95,
	title = {Multi-Modal Volume Registration by Maximization of Mutual Information},
	author = "W. Wells and P. Viola and R. Kikinis",
	booktitle = "Proceedings of the Second International Symposium on Medical Robotics and Computer Assisted Surgery",
	year = "1995",
	publisher = "Wiley",
	pages = "55 -- 62"
}

@article{Bienenstock82,
       author = "Bienenstock, E.L. and L.N. Cooper and P.W. Munro",
        title = "Theory for the Development of Neuron Selectivity: Orientation Specificity and Binocular Interaction in Visual Cortex",
      journal =  jneusci,
       volume =  2,
         year =  1982
  }

		  
@InProceedings{Baum95,
  author = 	 "Eric B. Baum and Dan Boneh and Charled Garrett",
  title = 	 "Where Genetic Algorithms Excel",
  year =	 1995,
  booktitle =	 colt,
  organization = ACM,
  address =	 "New York",
  annote =	 "Learning Genetic Algorithms"
}




@InProceedings{Lang95,
  author = 	 "Kevin Lang",
  title = 	 "Hill Climbing Beats Genetic Search on a Boolean Circuit
                  Synthesis Problem of Koza's",
  year = 	 1995,
  booktitle = "Twelfth International Conference on Machine Learning"
}

@Article{rissanen78,
  author = 	 "J Rissanen",
  title = 	 "Modeling by Shortest Data Description",
  journal =	 "Automatica",
  year =	 1978,
  volume =	 14,
  pages =	 "465-471"
}


@phdthesis(demarcken96c,
	author = "Carl de Marcken",
	title = "Unsupervised Language Acquisition",
	year = 1996,
	school = "Massachusetts Institute of Technology",
	address = "Cambridge, Massachusetts")





@Misc{QBIC,
  title =	 "The IBM QBIC Project",
  howpublished = "Web: http://wwwqbic.almaden.ibm.com/"
}

@Misc{VIRAGE,
  title =	 "The Virage Project",
  howpublished = "Web: http://www.virage.com/"
}

@Article{Niblack93,
  author = 	 "V. Niblack and R. Barber and W. Equitz and
		  M. Flickner and  E. Glasman and D. Petkovic and
		  P. Yanker and  C. Faloutsos and G. Taubin",
  title = 	 "The QBIC project: querying images by content using color, texture, and shape",
  journal =	 "IS&T/SPIE 1993
International Symposium on Electronic Imaging: Science & Technology",
  year =	 1993,
  volume =	 1908,
  pages =	 "173-187"
}

@Article{Kelly95,
  author = 	 "M. Kelly and T. M. Cannon and D. R. Hush",
  title = 	 "Query by image example: the CANDID approach",
  journal =	 "SPIE Vol. 2420 Storage and Retrieval for Image and Video
Databases III",
  year =	 1995,
  pages =	 "238-248"
}



@Article{LeCun89,
  author = 	 "Y. Le Cun and B. Boser and J. S. Denker and
		  D. Henderson and  R. E. Howard and W. Hubbard and L. D. Jackel",
  title = 	 "Backpropogation applied to handwritten zip code recognition",
  journal =	 "Neural Computation",
  year =	 1989,
  volume =	 1,
  pages =	 "541-551"
}


@TechReport{Pentland95,
  author = 	 "A. Pentland and R. W. Picard and S. Sclaroff",
  title = 	 "Photobook: Content-based manipulation of image databases",
  institution =  "MIT Media Lab",
  year = 	 1995,
  number =	 255
}


@Article{Picard93,
  author = 	 "R. W. Picard and T. Kabir",
  title = 	 "Finding similar patterns in large image databases",
  journal =	 "ICASSP",
  year =	 1993,
  volume =	 "V",
  pages =	 "161-164"
}

@TechReport{Rao95,
  author = 	 "R. P. N. Rao and D.H. Ballard",
  title = 	 "Object indexing using an iconic sparse distributed memory",
  institution =  "University of Rochester",
  year = 	 1995,
  number =	 "TR-559"
}

@Article{Turk91,
  author = 	 "M. Turk and A. Pentland",
  title = 	 "Eigenfaces for recognition",
  journal =	 "Journal of Cognitive Neuroscience",
  year =	 1991,
  volume =	 3,
  number =	 1,
  pages =	 "71-86"
}





@InProceedings{SimAde96,
  author = 	 "Eero P Simoncelli and Edward H Adelson",
  title = 	 "Noise Removal via Bayesian Wavelet Coring",
  booktitle =	 "IEEE Third Int'l Conf on Image Processing",
  year =	 1996,
  organization = "IEEE",
  address =	 "Laussanne Switzerland",
  month =	 "September"
}

@inproceedings{SaundMoran,
  author = 	 "E. Saund and T. Moran",
  title = 	 "Perceptual Organization in an Interactive Sketch Editing Application",
  booktitle = iccv,
  publisher = "IEEE, Washington, DC",
  address = "Cambridge, MA",
  month = "June",
  year = 	 1995,
}

		  
@book{DudaHart,
	author = "R.O. Duda and P.E. Hart",
	title = "{P}attern {C}lassification and {S}cene {A}nalysis",
	publisher = "John Wiley and Sons",
	year = "1973"
	}

@InProceedings{stahovich-davis-shrobe-96,
  author = 	 "T. F. Stahovich and R. Davis and H. Shrobe",
  title = 	 "Generating Multiple New Designs from a Sketch",
  pages =	 "1022--29",
  booktitle =	 "Proceedings Thirteenth National Conference on Artificial Intelligence",
  year =	 1996,
  organization = aaai
}

		  
@Article{palumbo-srihari-96,
  author = 	 "P.W. Palumbo and S.N. Srihari",
  title = 	 "Postal Address Reading in Real Time",
  journal =	 "International Journal of Imaging Science and Technology",
  year =	 1996
}