Neurocomputing

PYGMALION - 2059

Keywords neural networks, signal processing, neural networks development environments, image processing, speech processing

Start Date: 02-FEB-89 / Duration: 24 months

[ contact / participants ]

Objectives and Approach

The aim of PYGMALION was to create an independent European technological base for the applications, algorithms and software aspects of neurocomputing. The project led to the coordination of research on connectionist computing techniques and provided the means for developing the necessary software tools for productive research.

The objectives are to:

• promote information exchange within the European research community
• develop a portable European Neural Network Specification Language and tools to stand as an interface between applications and emulation architectures
• demonstrate the potential of a neural approach through industrial applications, mainly of image processing and speech processing.

Progress and Results

• Environment: the high and intermediate languages (N and nC) have been specified. They are close to the C++ language. The development of compilers is underway. All the well-known neural models are already available in the library.
• Hardware: a first cascadable VLSI demonstrator has been produced and tested. This chip offers full connectivity, which allows the programming of the most commonly used neural networks models (BP, Hopfield, MLP).
• Low-level imaging: results of applications of neural models (BP-constrained) have proved robustness and good restoration for image compression. Applications for real-time HDTV compression are under study. Further work on image segmentation and texture is underway, and is building the basis for future image processing using neural networks.
• High-level imaging, speech recognition and acoustic signal classification: positive results have been obtained and demonstrated in these three areas, with some potential short-term commercial applications.

Exploitation

Thomson-CSF intends to commercialise the general-purpose neurocomputer to serve as a fast application development environment, and the software environment for the investigation, simulation and analysis of neural models.

Philips intends to implement a neural network simulator on a DOOM machine.

Standard Elektrik Lorenz intends to make speech recognisers, based on artificial neural network methodology for office environments and information processing, as an OEM product.

CSELT will exploit progress from the demonstrator projects through the other companies of the STET holding (of which it is the advanced research centre).

The academic and research partners will contribute to the diffusion of the technology.

CONTACT POINT

Dr Catherine Simon
THOMSON-CSF-DSE
Boulevard de Valmy 160
F - 92704 Colombes
tel: + 33/ 1-41303597
fax: + 33/ 1-41303636
telex: 616780 TCSF

Participants

THOMSON-CSF-DSE - F - C
CNRS-ENS - F - P
CSELT - I - P
IRIAC - F - P
PHILIPS-LEP - F - P
COMPUTER TECHNOLOGY INSTITUTE - GR - P
INESC - P - P
STANDARD ELEKTRIK LORENZ AG - D - P
UNIVERSITY COLLEGE LONDON - UK - P
UNIVERSIDAD POLITECNICA DE MADRID - E - P
SEXTANT AVIONIQUE - F - A
POLITECNICO DI TORINO - I - A
INPG - F - A
CONCEPT LOGICIELS EXPERT - F - A

PYGMALION - 2059, December 1993

please address enquiries to the ESPRIT Information Desk

html version of synopsis by Nick Cook