A Remark on Alan H. Kawamoto: Nonlinear Dynamics in the Resolution of Lexical Ambiguity: A Parallel Distributed Processing Account
Reginald Ferber · 1993
Kawamoto (1993) reports a simulation of word recognition processes with a recurrent network of 216 units. 48 patterns of activity were used to train the network and yield data which are compared with data from word recognition experiments. Beside the presentation of his interesting data, Kawamoto describes his network as a dynamical system, sharing properties with a Hopfield net such as nonincreasing energy and convergence toward stable states. This note gives counter examples to some of the statements made in this description: Simple nets with increasing energy and not converging to stable states. It shows why the notion of “local energy minima ” is not of much use for networks updated in parallel and points out that there are important differences between the two types of networks. A simulation shows the strong influence of the representation of words Kawamoto uses. In his article ”Nonlinear Dynamics in the Resolution of Lexical Ambiguity: A Parallel Dis-tributed Processing Account ” Alan H. Kawamoto (1993) describes a simulation of word recog-nition processes with a recurrent network of 216 units. patterns of activity were used to train the network and yielded data which are compared to data from word recognition experiments.