Abordagem para redução de complexidade de RNA usando reconfiguração dinâmica.

Abstract

In this thesis a new solution for the treatment of the complexity in the interconnections among the processing elements of the artificial neural networks (ANNs) is described. It enables realize ANNs digital hardware implementation with a larger number of neurons than does nowadays. The ANNs have been used as a solution in various complex problems. Some of these problems require hardware implementation. A lot of constraints must be satisfied during the project flow of the implementations of ANNs, such as the neural interconnections. Nowadays, neural implementations are done using integrated circuits, specifically developed for a given neural network architecture or integrated circuits configured by the user. Among these circuits exist the dynamically reconfigured FPGAs (DR-FPGAs) which can have their characteristics changed during operation without suffering interruptions in their execution. These devices have been usedforANNimplementations. Itpresentsaproposaltosolvethe interconnection problem for artificial neurons using DR-FPGAs in a new computational way: the Execution Patterns1 (EPs). The EPs allow, theoretically, to reduce the influence of interconnections through the removal of data transport via busses, besides other advantages and disadvantages. TheEPsdoesnotseemtoberestrictedonlytoANNapplications. They can be used by reconfigurable computation in massive parallel problems and/or problems that demand information exchange among the various elements in a processing system.