Accuracy versus complexity in RBF neural networks

We have introduced a methodology for solving the tradeoff between accuracy and complexity in complex virtual systems directly at the system level. Such methodology can be inserted in an application-level compiler for transforming a high-level description of the application into a lower level. This can then be fed into a hardware/software codesign compiler for final system implementation. Off-line integration of constraints relaxes model accuracy by introducing the concept of neural networks equivalent according to accuracy. The complexity criterion help select the smallest neural network topology within this set. The methodology is the key element for an effective high-level synthesis where few or no application tuning parameters need to be set by the designer. Indirectly, the methodology supports a system level integration of the requirement for low-power consumption, a particularly appealing constraint for embedded system design