Identification of a specific limitation on local-feedback recurrent networks acting as Mealy-Moore machines

• Published in: IEEE transactions on neural networks Vol. 10; no. 2; pp. 433 - 438
• Main Author: Kremer, S.C.
• Format: Journal Article
• Language: English
• Published: New York, NY IEEE 1999; Institute of Electrical and Electronics Engineers
• Subjects: Activation; Applied sciences; Architecture; Automata; Automation; Computation; Computer architecture; Computer networks; Councils; Electric, optical and optoelectronic circuits; Electronics; Exact sciences and technology; Gradient methods; Information science; Logistics; Mathematical analysis; Mathematical models; Networks; Neural networks; Neurofeedback; Robustness; State feedback; Finite state machine; Network architecture; Identification; Mealy-Moore machine; Neural network; Dynamical system; Learning algorithm; Computing machine; Recurrent neuronal network
• ISSN: 1045-9227
• DOI: 10.1109/72.750574

Describes a proof which identifies limitations of local-feedback recurrent networks (LFRNs) in representing Mealy-Moore machines. Specifically, the paper identifies a class of machines that are unrepresentable in these networks. While it is known that LFRNs cannot represent all automata, this is the first paper to show a specific class of unrepresentable automata. The paper goes beyond previous work in the area by addressing realistic activation functions, arbitrary network architectures, and more general computing machines.