A representational redescription method using competitive learning

In a previous simulation a network learned labeling a small set of stimuli in three input conditions (visual features, label, label + visual features), classifying them according to colour, function, object name. Results suggested that in different input conditions the network internal representation reflects the explicit semantic structure of stimuli. Evidence on the mediating role of linguistic label was suggested by cluster analysis: in the three input conditions a single object is represented very similarly but it has different representations in the label + features condition, depending on the label. One limit of this kind of models is that transferring acquired knowledge to other tasks would require network retraining. A second shortcoming is that this model allows only one level of knowledge representation. Empirical evidence shows that knowledge must be represented at different levels, ranging from implicit to full explicit symbolism. A similar view has been suggested by the “representational redescription” hypothesis (Clark & Karmiloff-Smith, 1993). In order to test how to meet these requirements, we augmented our model requiring the network to use the already acquired knowledge to extract the semantic structure of the stimulus set, for each of the three subtasks. The hidden unit layer was connected to a new module with three clusters of output units. Each output cluster had to make explicit the structure in each of the three categorization subtasks. To train this new module, all the other connection weights were frozen except the connection from the hidden layer to the new output units. These connection weights were trained with the competitive learning algorithm. The results show that the network is able to exploit previously acquired knowledge and to make explicit the stimuli semantic structure using the hidden (implicit) representation. This structure corresponds to that obtained by cluster analysis in the previous research. This method can be considered a first step in testing the representational redescription hypothesis. It will require further exploration and testing of more complete models. Some related issues are discussed, such as the controversial need of hybrid models.