The synthesis of compact fuzzy neural circuits

Concerns Smart Parts, a class of fuzzy neural hardware. Smart Parts are "smart" in that they can learn an input-output mapping implicit in a data set. They are "parts" in that they are small high-speed fuzzy neural processors meant to provide the fuzzy hardware designer customized functionality in a small package. They are application-specific. The paper focuses on the tool assisting this synthesis, TROUT. Using design heuristics that favor small size and high speed, TROUT can produce Smart Parts circuit specifications virtually automatically. It starts by minimizing the complexity of the target application data set in a way that simplifies the eventual implementation. It chooses a neural net or fuzzy network model from a small library that best suits the data set. In this paper, we detail the fuzzy min-max classifier model (FMM). TROUT optimizes FMM learning parameters to produce the smallest circuit offering the highest input vector throughout. Two architectural insights make the synthesis tractable. First, the FMM network architecture is structurally adaptive. Second, asynchronous circuit-design techniques are used because they simplify the synthesis process by eliminating clock scheduling. The output from TROUT is very high-level hardware description language (VHDL) code that can be used to synthesize the circuit in any of a number of circuit technologies. We outline the synthesis process and provide a circuit example based on the public domain wine classification data set.