Real-time unsupervised neural networks for non-implementable in natural noise: A refutable hypothesis based on experiment

Summary form only given. An analog computer implementation of an unsupervised neural network was investigated. Results indicate that the implementation of the nonlinear ordinary differential equations of the Outstar learning model are destabilized by unavoidable multiplicative biases produced by physical circuitry, attributable to 1/f noise drift. A multiplicative offset perturbation model was developed to simulate the instabilities discovered in the analog implementation. Timing and scaling optimizations are required to allow stable learning of spatial patterns. These results were generalized with the hypothesis that real-time unsupervised neural networks are nonimplementable in natural noise. The perturbation model facilitates the testing of this hypothesis. It seems that 1/f phenomena are teleologically related to physically occurring self-organizing systems. The authors suggest a fractional calculus. Outstar generalization for incorporating nonzero mean noise-induced parameters caused by multiple scales of self-organizing interactions