CAVIAR: Convolution AER vision architecture for real-time

• Unité de recherche: PCRD EU
• Numéro de projet: 02.0273
• Titre du projet anglais: CAVIAR: Convolution AER vision architecture for real-time

Mots-clé (Anglais)

Forecasting; Information Processing; Information Systems; Innovation; Technology Transfer; Life Sciences

Autre Numéro de projet (Anglais)

EU project number: IST-2001-34124

Programme de recherche (Anglais)

EU-programme: 5. Frame Research Programme - 1.2.8 Generic R&D activities

Description succincte (Anglais)

See abstract

Résumé des résultats (Abstract) (Anglais)

Biological brains are structured in layers of neurons, where neurons in a layer connect to a very large number of neurons in the following layer. Each neuron in a layer connects to a projective field in the next layer. This can be approximated by two-dimensional convolutions. For Real-Time solutions direct hardware implementations are required. But hardware engineers face a very strong barrier: the massive connectivity. The problem gets worst for multi-chip multi-layer hierarchical bio-like systems. Address-Event-Representation (AER) is an incipient bio-inspired spike-based technique capable of providing a hardware solution.

The objectives of CAVIAR are two-fold:
1) To develop a general AER infrastructure for constructing bio-inspired hierarchically structured multi-chip systems for sensing + processing + actuation;
2) The implementation of a particular perceptive-action demonstrator vision system exploiting this infrastructure.

Objectives:
The objective of CAVIAR is to develop a robust AER infrastructure capable of supporting bio-inspired multi-chip multi-layer hierarchical sensing/processing/actuation systems. This will provide a unique platform to facilitate long-term research on and development of complex bio-inspired systems and to test novel ideas in spike-based processing. The following parts will be developed within the CAVIAR project: VLSI Chips: Sensing Retina Chip, Programmable-kernel Convolution Processing Chip, Dimension Reduction Competition 'Object' Chip, Spatio-temporal Learning Chip. AER Interfaces: Chip-to-Chip Interfacing Module, Chip-to-Computer Interfacing Module. Demonstrator: build a vision system for detecting balls and mount it on a robot. Bus Interactions and Consortium Standards: We will define a minimum set of AER standards that will let partners develop parts independently, such that it will be possible to interconnect them when assembling more complicated, hierarchical systems.

Work description:
The work is structured in 7 WPs:
WP1: Sensing Retina Chip. A medium-resolution retina chip will be provided that responds to local positive and negative irradiance transients in a two-dimensional input image. It performs focal-plane brightness adaptation and data compression for efficient use of the sensing and communication bandwidths;
WP2: Programmable-kernel Convolution Processing Chip. A processing chip will be provided that implements spatial filtering operations such as those found in biological neural systems. This chip will execute real-time convolution filtering operations on a two-dimensional data set. The convolution kernel will be programmable, such that a set of identical convolution chips can be used to extract different features at different spatial scales in parallel from a single data set;
WP3: Dimension Reduction Competition 'Object' Chip. We will develop a dimensionality-reduction and in-layer competition chip. It will combine several sparsely-coded two-dimensional maps, such as those provided by a set of convolution chips, select the dominant feature and output the coordinates of its centre of mass;
WP4: Spatio-temporal Learning Chip. A spatio-temporal pattern learning chip will be develop that performs unsupervised learning to recognize specific shapes, speeds and directions using a spike-based learning rule. This learning will help to predict the course of typical pattern sequences;
WP5: AER Interfaces. A set of chip-computer and chip-chip interfaces will be developed for testing, diagnosis, and development of hierarchical AER systems;
WP6: Demonstrator. To drive properly the development of the previous components the consortium will develop as well a demonstrator vision system in which all developed components will be put together on a robot platform for detecting and following a moving ball;
WP7: Bus Interactions and Consortium Standards. A set of common rules will be set, tested, and discusses for assembling AER components.

Références bases de données (Anglais)

Swiss Database: Euro-DB of the
State Secretariat for Education and Research
Hallwylstrasse 4
CH-3003 Berne, Switzerland
Tel. [phone]
Swiss Project-Number: 02.0273