SYSTOLIC ARRAY METHODOLOGY FOR A NEURAL MODEL TO SOLVE THE MIXTURE PROBLEM
R. Pérez, Pablo Martı́nez, Antonio Plaza, P. L. Aguilar · Series in machine perception and artificial intelligence · 2002
In this chapter we present a robust method for determining and quantifying components in a composite spectrum; we assume that the patterns of the individual spectra are known in advance. The proposed method is supported by a linear recurrent neural network based on the Hopfield model (HRNN). The HRNN has very important characteristics in the implementation of lowcomplexity VLSI structures, since only multiplying and adding operations are required for an inversion process. We propose a systolic array for its implementation. In order to describe the systolic structure, we use a methodology based on the dependence graph. We suggest a sequential algorithm that verifies the unique assignment rule and the locality of the data dependences. The proposed systolic structure is divided into two parts, one rectangular array for implementing the weight matrix determination, and a linear array for obtaining the Threshold vector. Both structures are used for realising the iterative process of the Neural Network. These structures allow us to reduce the computational cost from o(N2) to o(N) so that the weight matrix is obtained, and o(N3) to o(N) in order to realise the iterative process.