Path: utzoo!utgpu!utstat!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!ucbvax!HPLABS.HP.COM!neuron-request From: neuron-request@HPLABS.HP.COM ("Neuron-Digest Moderator Peter Marvit") Newsgroups: comp.ai.neural-nets Subject: Neuron Digest V5 #22 Message-ID: <18708.611097364@hplpm> Date: 13 May 89 21:16:04 GMT Sender: daemon@ucbvax.BERKELEY.EDU Reply-To: "Neuron-Request" Organization: Hewlett-Packard Laboratories Lines: 669 Neuron Digest Saturday, 13 May 1989 Volume 5 : Issue 22 Today's Topics: Preprints of two recent publications are available TR available research reports available CVPR89 Announcement TR: Virtual Memories and Massive Generalization POST-DOC: SPEECH & NEURAL NETS Peripheral N.S. and Homeostasis: BBS Call for Commentators ERPs, Memory and Attention: BBS call for Commentators Neural Networks Symposium Vision and Image Analysis Send submissions, questions, address maintenance and requests for old issues to "neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request" ARPANET users can get old issues via ftp from hplpm.hpl.hp.com (15.255.16.205). ------------------------------------------------------------ Subject: Preprints of two recent publications are available From: Jose Ambros-ingerson Organization: University of California, Irvine - Dept of ICS Date: 31 Mar 89 07:40:22 +0000 Preprints of two recent publications are available from the Computational Neuroscience Program at the University of California at Irvine: ================================================================== DERIVATION OF ENCODING CHARACTERISTICS OF LAYER II CEREBRAL CORTEX Richard Granger, Jose Ambros-Ingerson, and Gary Lynch Center for the Neurobiology of Learning and Memory University of California Irvine, CA. 92717 Computer simulations of layers I and II of piriform (olfactory) cortex indicate that this biological network can generate a series of distinct output responses to individual stimuli, such that different responses encode different levels of information about a stimulus. In particular, after learning a set of stimuli modeled after distinct groups of odors, the simulated network's initial response to a cue indicates only its group or category, whereas subsequent responses to the same stimulus successively subdivide the group into increasingly specific encodings of the individual cue. These sequences of responses amount to an automated organization of perceptual memories according to both their similarites and differences, facilitating transfer of learned information to novel stimuli without loss of specific information about exceptions. Human recognition performance robustly exhibits multiple levels: a given object can be identified as a vehicle, as an automobile, or as a Mustang. The findings reported here suggest that a function as apparently complex as hierarchical recognition memory, which seems suggestive of higher `cognitive' processes, may be a fundamental intrinsic property of the operation of this single cortical cell layer in response to naturally-occurring inputs to the structure. We offer the hypothesis that the network function of superficial cerebral cortical layers may simultaneously acquire and hierarchically organize information about the similarities and differences among perceived stimuli. Experimental manipulation of the simulation has generated hypotheses of direct links between the values of specific biological features and particular attributes of behavior, generating testable physiological and behavioral predictions. (Appears in Journal of Cognitive Neuroscience, 1:61-84, 1989). ==================================================== MEMORIAL OPERATION OF MULTIPLE, INTERACTING SIMULATED BRAIN STRUCTURES Richard Granger, Jose Ambros-Ingerson, Ursula Staubli and Gary Lynch Center for the Neurobiology of Learning and Memory University of California Irvine, CA. 92717 Primary findings from simulations of the superficial layers of olfactory cortex have been that repeated sampling of stimuli has two major effects: first, multiple samples greatly increase the information capacity of a network compared to that for a single sample, and second, the breaking of the response into distinct samples imposes an organization on the memories thus read out. It was found that repetitive sampling allows the network to form and read out a sequence of different representations of a stimulus, denoting information ranging from the membership of that stimulus in a group of similar stimuli, to specific information unique to the stimulus itself. This led us to the hypothesis that the combination of particular cellular physiological features, anatomical designs, and repetitive sampling performance, allows cortical networks to construct perceptual hierarchies (Lynch and Granger, 1989)*. Those initial simulation experiments did not address what is presumably an essential feature of repetitive sampling: namely, the interaction between the cortex and its inputs. The present paper reviews both our isolated cortical simulations and our first efforts to explore the issue of interaction between cortex and peripheral structures. New findings indicate that the mechanism of repeated sampling enables active analysis of stimuli into their learned components. *[Lynch, G. and Granger, R. (1989). Simulation and analysis of a cortical network. The Psychology of Learning and Motivation, Vol.23 (in press).] (To appear in: Neuroscience and Connectionist Models, M.Gluck and D.Rumelhart, Eds., Hillsdale: Erlbaum Associates, 1989.) =================================================== Send requests for reprints to: Richard Granger Computational Neuroscience Program Bonney Center University of California Irvine, California 92717 (granger@ics.uci.edu) ------------------------------ Subject: TR available From: "Jose del R. MILLAN" Date: 31 Mar 89 17:09:00 +0800 The following Tech. Report is available. Requests should be sent to MILLAN@FIB.UPC.ES ________________________________________________________________________ Learning by Back-Propagation: a Systolic Algorithm and its Transputer Implementation Technical Report LSI-89-15 Jose del R. MILLAN Dept. de Llenguatges i Sistemes Informatics Universitat Politecnica de Catalunya Pau BOFILL Dept. d'Arquitectura de Computadors Universitat Politecnica de Catalunya ABSTRACT In this paper we present a systolic algorithm for back-propagation, a supervised, iterative, gradient-descent, connectionist learning rule. The algorithm works on feedforward networks where connections can skip layers and fully exploits spatial and training parallelisms, which are inherent to back-propagation. Spatial parallelism arises during the propagation of activity ---forward--- and error ---backward--- for a particular input-output pair. On the other hand, when this computation is carried out simultaneously for all input-output pairs, training parallelism is obtained. In the spatial dimension, a single systolic ring carries out sequentially the three main steps of the learnng rule ---forward, backward and weight increments update. Furthermore, the same pattern of matrix delivery is used in both the forward and the backward passes. In this manner, the algorithm preserves the similarity of the forward and backward passes in the original model. The resulting systolic algorithm is dual with respect to the pattern of matrix delivery ---either columns or rows. Finally, an implementation of the systolic algorithm for the spatial dimension is derived, that uses a linear ring of Transputer processors. ------------------------------ Subject: research reports available From: Richard Zemel Date: Wed, 05 Apr 89 18:11:27 -0400 The following two technical reports are now available. The first report describes the main ideas of TRAFFIC. It appeared in the Proceedings of the 1988 Connectionist Summer School, Morgan Kaufmann Publishers, edited by D.S. Touretzky, G.E. Hinton, and T.J. Sejnowski. The second report is a revised version of my Master's thesis. It contains a thorough description of the model, as well as implementation details and some experimental results. This report is rather long (~75 pages), so if you are curious about the model we'll send you the first one. On the other hand, if you want to plough through the details, ask specifically for the second one. *************************************************************************** "TRAFFIC: A Model of Object Recognition Based On Transformations of Feature Instances" Richard S. Zemel, Michael C. Mozer, Geoffrey E. Hinton Department of Computer Science University of Toronto Technical report CRG-TR-88-7 (Sept. 1988) ABSTRACT Visual object recognition involves not only detecting the presence of salient features of objects, but ensuring that these features are in the appropriate relationships to one another. Recent connectionist models designed to recognize two-dimensional shapes independent of their orientation, position, and scale have primarily dealt with simple objects, and they have not represented structural relations of these objects in an efficient manner. A new model is proposed that takes advantage of the fact that given a rigid object, and a particular feature of that object, there is a fixed viewpoint-independent tranformation from the feature's reference frame to the object's. This fixed transformation can be expressed as a matrix multiplication that is efficiently implemented by a set of weights in a connectionist network. By using a hierarchy of these transformations, with increasing feature complexity in each successive layer, a network can recognize multiple objects in parallel. ****************************** "TRAFFIC: A Connectionist Model of Object Recognition" Richard S. Zemel Department of Computer Science University of Toronto Technical report CRG-TR-89-2 (March 1989) ABSTRACT Recent connectionist models designed to recognize two-dimensional shapes independent of their orientation, position, and scale have not represented structural relations of the objects in an efficient manner. A new model is described that takes advantage of the fact that given a rigid object, and a particular feature of that object, there is a fixed viewpoint-independent transformation from the feature's reference frame to the object's. This fixed transformation can be expressed as a matrix multiplication that is efficiently implemented by a set of weights in a connectionist network. The model, called TRAFFIC (a loose acronym for ``transforming feature instances''), uses a hierarchy of these transformations, with increasing feature complexity in each successive layer, in order to recognize multiple objects in parallel. An implementation of TRAFFIC is described, along with experimental results demonstrating the network's ability to recognize constellations of stars in a viewpoint-independent manner. ************************************************************************* Copies of either report can be obtained by sending an email request to: INTERNET: carol@ai.toronto.edu UUCP: uunet!utai!carol BITNET: carol@utorgpu ------------------------------ Subject: CVPR89 Announcement From: "Worthy N. Martin" Organization: U.Va. CS Department, Charlottesville, VA Date: 09 Apr 89 22:28:12 +0000 IEEE Computer Society Conference on COMPUTER VISION AND PATTERN RECOGNITION Sheraton Grand Hotel San Diego, California June 4-8, 1989 General Chair Professor Rama Chellappa Department of EE-Systems University of Southern California Los Angeles, California 90089-0272 Program Co-Chairs Professor Worthy Martin Professor John Kender Dept. of Computer Science Dept. of Computer Science Thornton Hall Columbia University University of Virginia New York, New York 10027 Charlottesville, Virginia 22901 Program Committee Chris Brown Avi Kak Theo Pavlidis Allen Hansen Rangaswamy Kashyap Alex Pentland Robert Haralick Joseph Kearney Azriel Rosenfeld Ellen Hildreth Daryl Lawton Roger Tsai Anil Jain Martin Levine John Tsotsos Ramesh Jain David Lowe John Webb John Jarvis Gerard Medioni General Conference Sessions will be held June 6-8, 1989 Conference session topics include: -- Edge Detection -- Shape from _____ (Shading, Contour, ...) -- Feature Extraction -- Motion -- Morphology -- Neural Networks -- Range Data: Generation and Processing -- Image and Texture Segmentation -- Monocular, Polarization Cues -- Stereo -- Object Recognition -- Visual Navigation -- Preprocessing -- Applications of Computer Vision -- Vision Systems and Architectures Invited Speakers: June 6 June 7 June 8 Prof. J. Feldman Prof. V.S. Ramachandran Prof. M.A. Arbib ICSI, Berkeley Univ. Calif., San Diego Univ. of Southern Calif. Time, Space and Form Visual Perception in Schemas, Computer Vision in Computer Vision Humans and Machines and Neural Networks Tutorials June 4, am June 5, am June 5, pm 1. Morphology and 3. Robust Methods for 5. Analog Networks for Computer Vision Computer Vision Computer Vision: R.M. Haralick W. Forstner Theory and Applications 2. Intermediate and 4. Parallel Algorithms C. Koch Low Level Vision and Architectures for 6. Model Based Vision M.S. Trivedi Computer Vision W.E.L. Grimson V.K.P. Kumar The IEEE Computer Society will also hold a workshop entitled: Artificial Intelligence in Computer Vision June 5, 1989 General Chair: Professor Rama Chellappa Program Co-Chairs: Professor J.K. Aggarwal and Professor A. Rosenfeld Conference Registration (for CVPR and Tutorials) Conference Department CVPR IEEE Computer Society 1730 Massachusetts Ave Washington, D.C. 20036-1903 (202)371-1013 Fees, before May 8 CVPR - $200 (IEEE Members, includes proceedings and banquet) - $100 (Students, includes proceedings and banquet) Tutorials - $100 per session (IEEE Members and Students) Hotel Reservations Sheraton Grand Hotel on Harbor Island 1590 Harbor Island Drive San Diego, CA 92101 (619)692-2265 Rooms - $102 per night (single or double) The Advance Program with registration forms, etc. will be mailed out of the IEEE offices shortly. ------------------------------ Subject: TR: Virtual Memories and Massive Generalization From: Paul Smolensky Date: Thu, 13 Apr 89 09:45:53 -0600 Virtual Memories and Massive Generalization in Connectionist Combinatorial Learning Olivier Brousse & Paul Smolensky Department of Computer Science & Institute of Cognitive Science University of Colorado at Boulder We report a series of experiments on connectionist learning that addresses a particularly pressing set of objections on the plau- sibility of connectionist learning as a model of human learning. Connectionist models have typically suffered from rather severe problems of inadequate generalization (where generalizations are significantly fewer than training inputs) and interference of newly learned items with previously learned items. Taking a cue from the domains in which human learning dramatically overcomes such problems, we see that indeed connectionist learning can es- cape these problems in *combinatorially structured domains.* In the simple combinatorial domain of letter sequences, we find that a basic connectionist learning model trained on 50 6-letter se- quences can correctly generalize to over 10,000 novel sequences. We also discover that the model exhibits over 1,000,000 *virtual memories*: new items which, although not correctly generalized, can be learned in a few presentations while leaving performance on the previously learned items intact. Virtual memories can be thought of states which are not harmony maxima (energy minima) but which can become so with a few presentations, without in- terfering with existing harmony maxima. Like generalizations, virtual memories in combinatorial memories are largely novel com- binations of familiar subpatterns extracted from the contexts in which they appear in the training set. We conclude that, in com- binatorial domains like language, connectionist learning is not as harmful to the empiricist position as typical connectionist learning experiments might suggest. Submitted to the annual meeting of the Cognitive Science Society. Please send requests to conn_tech_report@boulder.Colorado.EDU and request report CU-CS-431-89. These will be available for mailing shortly. ------------------------------ Subject: POST-DOC: SPEECH & NEURAL NETS From: Ron Cole Date: Mon, 17 Apr 89 19:08:54 -0700 POST-DOCTORAL POSITION AT OREGON GRADUATE CENTER Speech Recognition with Neural Nets A post-doctoral position is available at the Oregon Graduate Center to study connectionist approaches to computer speech recognition, beginning Summer or Fall, 1989. The main requirements are (1) a strong background in the theory and application of neural networks, and (2) willingness to learn about the wonderful world of speech. Knowledge of computer speech recognition is helpful but not required; the PI has extensive experience in the area and is willing to teach the necessary skills. The goal of our research is to develop speech recognition algorithms that are motivated by research on hearing, acoustic phonetics and speech perception, and to compare performance of algorithms that use neural network classifiers to more traditional techniques. In the past year, our group has applied neural network classification to several problem areas in speaker-independent recognition of continuous speech: Pitch and formant tracking, segmentation, broad phonetic classification and fine phonetic discrimination. In addition, we have recently demonstrated the feasibility of using multi-layered networks to identify languages on the basis of their temporal characteristics (preprint available from vincew@ogc.cse.edu). OGC provides an excellent environment for research in speech recognition and neural networks, with state-of-the-art speech processing software (including Dick Lyon's cochleogram, a representation based on a computational model of the auditory system), speech databases, and simulation tools. The department has a Sequent Symmetry multiprocessor, Intel Hypercube and Cogent Research XTM parallel workstations, and the speech project has several dedicated Sun4 and Sun3 workstations. The speech group has close ties with Dan Hammerstrom's Cognitive Architecture Project at OGC, and with Les Atlas and his group at the University of Washington. OGC is located ten miles west of Portland on a spacious campus in the heart of Oregon's technology corridor. Nearby companies include Sequent, Intel, Tektronix, Cogent Research, Mentor Graphics, BiiN, NCUBE, and FPS Computing. The cultural attractions of Portland are close by, and the Columbia River Gorge, Oregon Coast and Cascade Mts (skiing through September) are less than 90 minutes away. Housing is inexpensive and quality of life is excellent. Please send resume to: Ronald Cole Computer Science and Engineering Oregon Graduate Center 19600 N.W. Von Neumann Drive Beaverton, OR 97006-1999 503 690 1159 ------------------------------ Subject: Peripheral N.S. and Homeostasis: BBS Call for Commentators From: harnad@Princeton.EDU (Stevan Harnad) Date: Fri, 21 Apr 89 23:12:06 -0400 Below is the abstract of a forthcoming target article to appear in Behavioral and Brain Sciences (BBS), an international, interdisciplinary journal that provides Open Peer Commentary on important and controversial current research in the biobehavioral and cognitive sciences. Commentators must be current BBS Associates or nominated by a current BBS Associate. To be considered as a commentator on this article, to suggest other appropriate commentators, or for information about how to become a BBS Associate, please send email to: harnad@confidence.princeton.edu or harnad@pucc.bitnet or write to: BBS, 20 Nassau Street, #240, Princeton NJ 08542 [tel: 609-921-7771] ____________________________________________________________________ BETA-AFFERENTS: A FUNDAMENTAL DIVISION OF THE NERVOUS SYSTEM MEDIATING HOMEOSTASIS? James C. Prechtl & Terry L. Powley Laboratory of Regulatory Psychobiology Department of Physiological Sciences Purdue University West Lafayette, IN 47907 Keywords: autonomic nervous system; capsaicin; dorsal root ganglion; neuroimmunology, nerve growth factor; nociception; sypathetics; substance P; sensory neurons; tachykinins; visceral afferents The peripheral nervous system (PNS) has classically been subdivided into a somatic division composed of both afferent and efferent pathways and an autonomic division containing only efferents. Langley, who codified this asymmetrical plan at the beginning of the 20th century, considered different afferents, including visceral ones, as candidates for inclusion in his concept of the "autonomic nervous system" (ANS), but he finally excluded all candidates for lack of any distinguishing histological markers. Langley's classification has been enormously influential in shaping modern ideas about both the structure and the function of the PNS. Here we survey modern information about the PNS and argue that many of the sensory neurons designated as "visceral" and "somatic" are in fact part of a histologically distinct group of afferents dedicated to autonomic function. These afferents have traditionally been known as "small dark" neurons or B-neurons. In this target article we outline an association between autonomic and B-neurons based on ontogeny, cell phenotype and functional relations, grouping them together as part of a common reflex system involved in homeostasis. This more parsimonious classification of the PNS, provided by the identification of a group of afferents associated primarily with the ANS, avoids a number of confusions produced by the classical orientation. It may also have practical implications for our understanding of nociception, homeostatic reflexes and the evolution of the nervous system. ------------------------------ Subject: ERPs, Memory and Attention: BBS call for Commentators From: harnad@Princeton.EDU (Stevan Harnad) Date: Fri, 21 Apr 89 23:17:58 -0400 Below is the abstract of a forthcoming target article to appear in Behavioral and Brain Sciences (BBS), an international, interdisciplinary journal that provides Open Peer Commentary on important and controversial current research in the biobehavioral and cognitive sciences. Commentators must be current BBS Associates or nominated by a current BBS Associate. To be considered as a commentator on this article, to suggest other appropriate commentators, or for information about how to become a BBS Associate, please send email to: harnad@confidence.princeton.edu or write to: BBS, 20 Nassau Street, #240, Princeton NJ 08542 [tel: 609-921-7771] ____________________________________________________________________ Keywords: selective attention, echoic memory, cortical localization, audition, orienting response, automatic processing THE ROLE OF ATTENTION IN AUDITORY INFORMATION PROCESSING AS REVEALED BY EVENT-RELATED POTENTIALS Risto Naatanen Department of Psychology University of Helsinki Helsinki, Finland This target article examines the roles of attention and automaticity in auditory processing as revealed by event-related potential (ERP) research. An ERP component called the "mismatch negativity" indicates that physical and temporal features of auditory stimuli are fully processed whether or not they are attended. It also suggests that there exists a mechanism of passive attention switching with changes in repetitive input. ERPs also reveal some of the cerebral mechanisms by which acoustic stimulus events produce and control conscious perception. The "processing negativity" component implicates a mechanism for attending selectively to stimuli defined by certain physical features. Stimulus selection occurs in the form of a matching process in which each input is compared to the "attentional trace," a voluntarily maintained representation of the task-relevant features of the stimulus to be attended. ------------------------------ Subject: Neural Networks Symposium From: ukrainec@maccs.McMaster.CA (Andy Ukrainec) Organization: McMaster U., Hamilton, Ont., Can. Date: Mon, 24 Apr 89 20:48:58 +0000 ----------------- | Neural Networks | ----------------- Communications Research Laboratory / TRIO One-Day Symposium June 26th, 1989 8:30 a.m. - 5:00 p.m. Pre-registration by mail: ($100.00) (Students - $25.00) Canadian Deadline for registrations: June 1st, 1989 For further information contact: Anne Myers CRL, McMaster University 1280 Main St. West Hamilton, Ontario, Canada L8S 4K1 (416) 525-9140 ext. 4085 Guest Speakers: - Dr. Isabelle Guyon, AT&T Bell Labs - Mr. Gary Josin, Neural Systems Inc., B.C. - Dr. Paul J. Werbos, Nat. Sc. Foundation, Washington, D.C. - Dr. B. Widrow, Stanford University, California ! Andrew Ukrainec ukrainec@maccs.mcmaster.ca "I am what I am!" ! ! Communications Research Laboratory, McMaster University ! ! Hamilton, Ontario, Canada L8S 4K1 ! ------------------------------ Subject: Vision and Image Analysis From: j daugman Date: Tue, 02 May 89 10:56:56 -0400 Request for Technical Reports and Papers (Second Request) In preparation for upcoming Reviews and Tutorials at 1989 Conferences, I would be grateful to receive copies of any papers or technical reports pertaining to applications of neural nets to vision and image analysis. (This repeats an earlier request sent out in February.) Please send any material to the following address. Thank you in advance. John Daugman 950 William James Hall Harvard University Cambridge, Mass. 02138 ------------------------------ End of Neurons Digest *********************