An artificial neural network-based codebook search technique

Abstract

An unsupervised learning artificial neural network (ANN) [T. Kohonen, Helsinki University of Technology Report TKK-F-A601 (1986)] was modified for a vector-quantized PCM (VPCM) codebook search problem. The ANN was structured such that each element of the network had associated with it a single codebook vector. The amount of processing required at each element of the network was used to derive an upper bound on the number of network iterations. Simulations were performed to determine the effect that various network parameters had on the speed of network convergence. The parameter values offering the greatest performance were applied to the network. The speed and computational complexity of the ANN solution to this problem were then compared to the same criteria for a standard linear (full-codebook) search technique. Analysis and test results indicate that the ANN approach can provide the speed of a tree search coupled with the minimum memory characteristics associated with a linear search, at the expense of requiring a multiprocessor configuration. [Work supported by NSF.]