D. K. Naik
R. J. Mammone
Rutgers Univ., CAIP Ctr., Core Bldg., Frelinghuysen Rd., Piscataway, NJ 08855
A new approach to adaptive language acquisition is presented. Previous methods for language acquisition have used an information theoretic approach to determine the connection between words and actions. A novel approach to language acquisition was recently developed, where a linear neural network is adaptively trained to map a set of words into a specified set of actions [A. Gorin et al., ``Adaptive Acquisition of Language,'' in Neural Networks, Theory and Applications, edited by R. J. Mammone and Y. Y. Zeevi (Academic, New York), p. 125]. This training can be used as a knowledge source to guide a neural network in acquiring associations more quickly in learning related tasks. This generalization of learning one task to help guide the learning of a related task is called metaphoric learning or learning by learning. A new method for metaphoric learning in neural networks is presented. The method uses an observing neural network called a meta-neural network (MNN) to direct the training of a conventional word-to-action mapping neural network. The MNN provides the conventional neural network with gradient information that is based on successful training strategies acquired previously. If the conventional neural network is exposed to words that are similar to those learned previously, the MNN is shown to improve the learning rate. The proposed scheme can also be applied to the problem of learning of different languages that are structurally similar to a language learned previously. Computer simulation results are presented that demonstrate improved learning rates for related language acquisition tasks.