On Training Neural Network Algorithms for Odor Identification for Future Multimedia Communication Systems

Abstract

Future multimedia communication system can be developed to identify, transmit and provide odors besides voice and image. In this paper, an improved odor identification method is introduced. We present an analysis of centergradient and a new method of using convergence parameters in training RBFN-SVD-SG (Radial Basis Function Network using Singular Value Decomposition combined with Stochastic Gradient) algorithm for odor identification. Through mathematical analysis, it was found that the steady-state weight fluctuation and large values of convergence parameter can lead to an increase of variance of center-gradient, which induces ill-behaving convergence. The proposed method of using raised-cosine functions for time-decaying convergence parameter shows faster convergence and better recognition performance.

Signal Processing Techniques Based on Adaptive Radial Basis Function Networks for Chemical Sensor Arrays

“…The use of a time-varying convergence parameter can be a solution to the problem of the RBF-SG algorithm. In [9] a modified version of a raised-cosine function is used as a time-decreasing function and is applied to a convergence coefficient for center adaptation.…”