Constructive learning of recurrent neural networks: Limitations of recurrent cascade correlation and a simple solution - Comment

Giles et al. have proven that Fahlman's recurrent cascade correlation (RCC) architecture is not capable of realizing finite state automata that have state-cycles of length more than two under a constant input signal. This paper extends the conclusions of Giles et al. by showing that there exists a corollary to their original proof which identifies a large second class of automata, that is also unrepresentable by RCC.