|
NeuroCOLT
Technical Report NC-TR-99-047
Dependencies
of composite connections in Field Programmable Neural Arrays
Bernard Girau,
LORIA INRIA-Lorraine
F-54506 Vandoeuvre-les-Nancy
Abstract
Neural networks are considered as naturally parallel
computing models. But the number of operators and the complex connection
graph of standard neural models can not be handled by digital hardware
devices. Neural network hardware implementations have to reconcile
simple hardware topologies with complex neural architectures. This
may be performed by means of some programmable hardware principles
applied to neural computation. Yet the definition of such original
neural models leads to complex weight dependencies. These models are
shortly described. They are analyzed along the underparameterized
convolution terms they compute. Theoretical points of view help to
estimate the unfavourable consequences of this underparameterization,
whereas applications to classification problems balance this estimation.
Good performances are reached with these models despite simplified
topologies. Learning takes advantage of the preceding theoretical
approach.
Download Compressed Postscript
|