Cognitio 2011

Abstract: In psychology, the issue of modeling the human categorization process has been ongoing for the better part of 50 years. Over the years many models have been proposed. One Particularly interesting example is the COVIS dual model of categorization [1]. COVIS posits the existence of two distinct systems of categorization, namely an implicit system responsible for learning non-verbalizable rules and an explicit system responsible for learning verbalizable rules.  Often in current literature models based around artificial neural networks are used to represent human cognitive processes as a distributed process using a large number of connected individual units. The resemblances between this architecture and what is known of neural networks in the brain permit what we believe is a faithful, if simplified, representation of basic cognition. Ashby used a connexionnist model composed of two neural networks to represent the human cognitive process through a computational representation. The first network is a Perceptron-like network that models the implicit system. The second network uses a rule-selection algorithm that models the explicit system. Finally, competition between the two determines which system is used. We wanted to take advantage of the property of self-organising maps that permits similar items to be represented topologically near one another to create the areas of mental space separated by the decision bound, as proposed by Ashby, Alfonso-Reese, Turken & Waldron (1998). In this study, it is proposed an implementation of the COVIS model using Feature-Extracting Bidirectional Associative Memories coupled with Self-Organising Maps (FEBAM-SOM). This model was developed to show unsupervised learning of perceptual patterns (through input compression), the concurrent encoding of proximities in a multidimensional space and the reconstruction of perfect outputs from incomplete and noisy patterns (Chartier, Giguère & Langlois, 2009). . This type of network allows the development of attractors centered on the k highest activated units surrounded by a neighbourhood function, representing the areas of mental space. In the proposed implementation, the implicit system is a FEBAM-SOM that process the whole stimulus, giving the same importance to all features. The explicit system decomposes the stimulus into its features and starts the categorization with a simple rule (1 feature) and elaborates it in function of its performance. The rule-choosing criterion is a ratio of the intra-category resemblance over the inter-category resemblance. Each of the rules and the implicit system all iterated the same fixed amount of times. Finally a comparison of this same ratio present in the final categorization of each system determines which system is to be used. The model was tested on  3 sets of 24 stimuli composed of 5 features, belonging to one of two categories. Each set represent one of three different types of categorization task; namely an information integration task, a conjunctive rule-based task or a disjunctive rule-based task. The stimuli are 5 x 10 x 10 tensors with continuous values ranging from -1 to 1. Results show that effective categorization were obtained for all 3 tasks with the model with the creation of an attractor for each category and the use of each system as predicted by the hypothesis, namely the implicit system for the information integration task and the explicit system for each of the rule-based tasks.