1D and 2D systolic implementations for radial basis function networks

Abstract

In this paper]. we show that Radzal Basis Function networks can be efficiently implemented on 1Dand 2 0 systolic arrays. We dascuss such networks in the framework of probability density function approximation f o r classification problems. In fact, the most computation intensive parts of the classification process consist an calculating pattern distances. In the znitaalzsataon phase of the algorithm this involves calculatang the mutual (antra-class) distance matrix. The classafication of an anput vector mainly involves the calculataon of the distance vector between this input vector and all the learnang set vectors.The proposed implementations are 2 0 systolic, 1D pipeline and 10 parallel. Practical implementation issues are discussed for the M A N T R A machine ( 2 0 grid) and the SMART Neuro-computer (1D ring).

Dynamically reconfigurable systolic array accelerators: a case study with extended Kalman filter and discrete wavelet transform algorithms

“…A subset of these algorithms and their SA implementations include: DWT [26], K-means clustering [27], Bayes classifier [28], eigenvalue calculation [29], etc. The DWT algorithm was selected for this application because it is a powerful filtering algorithm that has been used in aerospace applications [30,31] image compression algorithm can now be applied to the result of the DWT and achieve very efficient compression, due to the sparsity created by the DWT [31].…”