Skip to Main content Skip to Navigation
Theses

Contribution à la réalisation électronique de réseaux de neurones formels : intégration analogique d'une Machine De Boltzmann

Abstract : The integration capabilities of MOS technologies has been doubling every eighteen months for twenty years. An attractive approach to exploit this progress has been recently introduced through the invention of “analog cellular circuits” which are built out of the regular interconnexion of thousands identical analog cells. This Thesis is a contribution to the design knowledge of such circuits. It studies those which can be used in pattern recognition applications and which can be modelled by statistical Artificial Neural Networks (ANN). It focuses on the Boltzmann Machine, which is Boolean, stochastic and discrete time ANN model. In Chapter I, the Artificial Neural Network and the algorithms of Synchronous Boltzmann Machine are presented. The Synchronous model is well suited to a parallel implementation. In Chapter II, the State of the art of Boltzmann Machines realizations and the analog techniques used for ANN implementation are presented. In Chapter III, an original architecture of circuits dedicated to multi-layered networks executing Boltzmann Machine is presented. In Chapter IV, the design and simulations of the cells and circuits are described : voltage to current and current to voltage converters, comparator, sigmoidal function. This ANN model requires a stochastic implementation, so two original random generators are presented : the first one is opto-electronic and it takes advantage of the optical speckle properties, the second one is electronic and it uses a chaotic cellular automata. Some measures which have been done on prototype circuits are presented in Chapter V. In summary, future evolution of these techniques towards continuous time implementation and analog neurons state are described.
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00008989
Contributor : Eric Belhaire <>
Submitted on : Monday, April 11, 2005 - 11:30:40 AM
Last modification on : Wednesday, September 16, 2020 - 4:54:16 PM
Long-term archiving on: : Friday, April 2, 2010 - 9:39:22 PM

Identifiers

  • HAL Id : tel-00008989, version 1

Collections

Citation

Eric Belhaire. Contribution à la réalisation électronique de réseaux de neurones formels : intégration analogique d'une Machine De Boltzmann. domain_stic.othe. Université Paris Sud - Paris XI, 1992. Français. ⟨tel-00008989⟩

Share

Metrics

Record views

417

Files downloads

1666