home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.ai.neural-nets:4707 sci.physics:22131 sci.image.processing:1491 triangle.talks:234
- Path: sparky!uunet!gatech!concert!borg.cs.unc.edu!not-for-mail
- From: marshall@cs.unc.edu (Jonathan Marshall)
- Newsgroups: comp.ai.neural-nets,sci.physics,sci.image.processing,image,triangle.talks
- Subject: Triangle NN talks: Szu
- Followup-To: comp.ai.neural-nets
- Date: 5 Jan 1993 17:03:21 -0500
- Organization: The University of North Carolina at Chapel Hill
- Lines: 106
- Distribution: usa
- Expires: 21 Jan 93
- Message-ID: <1id0j9INNku2@marshall.cs.unc.edu>
- NNTP-Posting-Host: marshall.cs.unc.edu
-
- From: <sutton@eos.ncsu.edu>
-
- ============= TRIANGLE AREA NEURAL NETWORK SOCIETY presents: =============
-
- The Triangle Area Neural Network Society proudly presents Dr. Harold Szu,
- President of the International Neural Network Society, for two lectures on
- January 20, 1993.
-
- The first lecture will be at NCSU (and will be broadcast over the MCNC
- Network to requesting sites) at 10:15. The second will be at the Research
- Triangle Institute at 5:15, and refreshments will be available at 5 pm.
- These lectures are appropriate for all those with a neural network interest.
-
- ============================================================================
-
- The first lecture:
-
- Topic: REVIEW OF NEUROCOMPUTING, THE PAST AND THE FUTURE
- Speaker: Dr. HAROLD SZU
- Date/Time: January 20 1993, 10:15 to 11:05 AM
- Location: Studio 2, Park Shops, North Carolina State University
-
- Abstract:
- "Classical neural networks (NN) will be characterized with Hebbian-like
- learning algebra (small perturbation) and various architectures (fixed
- layers). The limited ability of fault tolerance, associative memory
- capacity, and the tradeoff between abstract/generalization and network
- complexity has now become obvious rooted in the small perturbation learning
- approach to a fixed architecture and the input impedance mismatch of sensor
- data format with the brain-style computing.
- "Then, the computational property of Biological NN (BNN) is observed in
- vitro by an electronic chip substrate system that consists of time-lapsed
- Video imaging (to be shown during the presentation) through a microscope
- (and some neurochemical control of synaptic growth, namely the phosphopro-
- teins: Synapsin IIb discovered by Han & Greengard, Nature 1991 could be
- added). The purpose of this on-going study is threefold:
- (1) Near Term: to measure in parallel with modern instrumentations the
- sigmoidal function proposed by McCullouch-Pitts five decades ago for a
- single neuron firing rate transfer function.
- (2) Middle Term: to characterize the neurite formation pair correlation
- function proposed recently.
- (3) Long Term: to determine in vitro the link between the specific BNN
- architectural topology to the functional efficiency in information
- processing. The minimum intelligent network is discovered to be
- three hairy neurons of chick embryo neurons, called Peter, Paul, Mary
- revealing in a dynamic reconfiguration in network architecture.
- "It should not be expected form our effort to unveil human intelligence,
- but only to capture and endow artificial neural networks (ANN) with somewhat
- intelligent capability for a devoted and efficient processing system."
-
- ============================================================================
-
- The second lecture:
-
- Topic: ADAPTIVE WAVELET TRANSFORM
- Speaker: Dr. HAROLD SZU
- Date/Time: January 20, 1993, 5:15 to 6:15 PM
- Location: Dreyfus Auditorium, Research Triangle Institute
-
- Abstract:
- "During the last decade, 1980's, French geologists & scientists, Morlet,
- Meyer, Grossmann, et. al., have developed the Wavelet Transform (WT) of a
- constant resolution resonator Q = df/f, in order to overcome the high
- frequency noisy seismic signal reconstruction instability that a fixed
- windowed Gabor Transform based on the uncertainty principle seems to
- produce. WT was demonstrated ideally for Wideband Transient (WT) signal
- syntheses, such as the sound, seismic wave image reconstruction for oil
- exploration. The major contribution of WT is the mathematics methodology to
- construct in the Hilbert space a complete and orthonormal (CON) set of
- daughter wavelets based on an affine transform of a good mother h(t) that is
- a band-pass function (no d.c. component) and has a finite energy (square-
- integrable). The affine transform can produce a so-called snug frame
- (almost CON coined by Daubechies) for a good enough mother to spawn a set of
- daughter wavelets hab(t) = h(t') by replacing t by t' = t P b/a, where b is
- the discrete or continuous shift in time and a is the scale parameter
- proportional to the Fourier frequency f expressed usually in a discrete
- logarithmic 2-base: 1/f = a = 2N, with N = 0, 11, 12, etc.
- "One of the recent advancements in WT is in the joint areas of neural
- networks and WT: `Neural Network Adaptive Wavelets for Signal Representation
- and Classification [Szu, et. al. Opt. Eng. 31, 1907Q1916, Sep.'92],' which
- is inspired by the sensitive human ear detection of one's own name or other
- interesting sound with P3dB during a noisy drinking party (Cocktail Party
- Effect). Another less obvious advancement in WT is perhaps in the shift of
- paradigm in solving a nonlinear dynamics problem `Why the soliton wavelet
- transform is useful for nonlinear dynamic phenomena,' Szu, H., SPIE 1705,
- 280Q288, 1992. While linear FT defers the nonlinearity to later mode-mode
- coupling, WT gives us the freedom to pay the nonlinear price early by
- choosing a particular nonlinear solution, say soliton, as a mother wavelet
- and then enjoy the linear superposition principle of WT."
-
- ============================================================================
-
- Dr. Szu is currently the Information Science Group Leader at Naval Surface
- Warfare Center, Code R44, White Oak, MD 20903. He is President of the
- International Neural Network Society and Adjunct Professor at George
- Washington University and American University.
-
- ============================================================================
-
- DIRECTIONS to Dreyfus Auditorium from Raleigh, Durham, or Chapel Hill: Take
- the Davis Drive Exit on I40. Turn right onto Davis Drive. At Cornwallis
- Road, which is the first intersection, go left. Take the first road on the
- left off Corwallis, which is Institute Drive. The second turn-off on the
- right leads directly to Dreyfus Auditorium.
-
- ============================================================================
-