Mapping across domains without feedback: A neural network model of transfer of implicit knowledge

Connectionist models of artificial grammar learning: what type of knowledge is acquired?

“…Besides the basic grammatical effect, the model has been shown to reproduce several additional effects such as the effects of similarity to training strings and grammaticality reported by Vokey and Brooks (1992;see Dienes, Altmann, & Gao, 1999) and effects of knowledge about the position of single letters (Kinder, 2000). A variant of the model has been demonstrated to explain transfer of grammar knowledge to a new letter set (e.g., Shanks, Johnstone, & Staggs, 1997; see Dienes, Altmann, & Gao, 1999). Furthermore, the model accounts for dissociations between classification and recognition performance in amnesic and control participants in the AGL paradigm (Kinder & Shanks, 2001).…”

The knowledge acquired during artificial grammar learning: Testing the predictions of two connectionist models

“…Because this accuracy is higher for grammatical than for nongrammatical strings, the model correctly predicts the former to be endorsed more frequently than the latter. Dienes (1992) compared several variants of the autoassociator model with respect to their capability of reproducing empirical data obtained from Dienes, Broadbent, and Berry (1991) as well as Dulany, Carlson, and Dewey (1984). The variants diered with respect to several features, these being the learning rule (Hebb rule vs. Delta rule), the coding of the letter features (single letter coding vs. additional coding of bigrams), the coding of absent features (activation of zero vs. activation of )1 if the letter is absent), successive versus simultaneous prediction (successive prediction: each unit received input only from units representing preceding letter positions; simultaneous prediction: each unit received input from all other units, except from itself), and the amount of training (preasymptotic training: the model was trained with the identical number of trials as were the participants; asymptotic training: the model was trained until asymptote was reached).…”

“…Several attempts have been made to investigate the kind of information which is stored in arti®cial grammar learning (Knowlton & Squire, 1994, 1996Perruchet & Pacteau, 1990;Shanks, Johnstone, & Staggs, 1997). Parallel to the growing interest in this issue of research, computational models of arti®cial grammar learning have been proposed (Dienes, 1992;Dienes, Altmann, & Gao, 1999;Servan-Schreiber & Anderson, 1990). It is crucial for the validity of these models that they make correct predictions about the kind of information which is stored during training and used for categorization.…”

“…However, they are still processed sequentially, since participants spell the strings from left to right in order to memorize them. Thus, the SRN model can also be applied to this kind of task (e.g., Dienes, Altmann, & Gao, 1999). Translation of the model's output into a response probability is basically the same as in the autoassociator: the letters predicted by the model are compared with the actual training string.…”

“…The models are dierent variants of an autoassociator and a simple recurrent network (SRN). The autoassociator is a single-layered model which consists of units, each representing a letter in a particular position (Dienes, 1992). Every unit is connected to every other one, but not to itself.…”