The Acquisition of New Categories through Grounded Symbols: An Extended Connectionist Model
چکیده
Solutions to the symbol grounding problem, in psychologically plausible cognitive models, have been based on hybrid connectionist/symbolic architectures, on robotic approaches and on connectionist only systems. This paper presents new simulations on the use of neural network architectures for the grounding of symbols on categories. In particular, the connectivity patterns between layers of the networks will be manipulated to scale up the performance of current connectionist models for the acquisition of higher-order categories via grounding transfer. 1 The Grounding of Symbols in Categories Cognitive models dealing with linguistic and symbol-manipulation tasks can use symbols that are either grounded or ungrounded (i.e. self-referential). Grounded symbols are those inherently significant to the cognitive system, such as an agent, and not mediated by the interpretation of an external user. Self-referential symbolic systems are those that use symbols that have no grounding in any other module of the cognitive agent. It has been claimed [5] that the cognitive relevance and psychological plausibility of a self-referential symbolic system is diminished as a result of the symbol grounding problem. To solve the problem, Harnad [5] suggested that symbols should be intrinsically linked to the agent’s ability of acquiring categories from everyday experience it has of its environment. In particular, it is necessary that some basic symbols are directly grounded on sensorimotor categories. Subsequently, new (grounded) categories can be formed through the combination of previously grounded basic symbols. Hybrid symbolic-connectionist models were originally proposed as ideal candidates for solving the symbol grounding problem [6]. More recently, alternative approaches have been introduced. Robotics approaches to symbol grounding focus on social learning and interaction between agents (including robots, internet agents and humans) to ground shared symbol communication systems. This has been implemented, for example, in experiments on robotic language games [9]. Fully connectionist models have also been proposed to deal with the symbol grounding problem [1,7,8]. For example, in [1] the ability of neural networks to acquire a small set of basic categories through direct sensorimotor grounding was tested. The same networks were subsequently trained to acquire new higher-order categories solely through combination of the name of basic categories (symbolic theft). These networks were able to transfer the grounding from sensorimotor categories to higherorder categories learnt via symbol combination. Such an approach has also been used in evolutionary simulations of language origins [2]. Research on the connectionist implementation of grounded symbolic cognitive agents is still in progress. In particular, effort has focused on the design of modular connectionist architectures and its contribution in dealing with the nature/nurture debate (e.g. [3]). This paper presents new simulations based on the manipulation of the connectivity pattern of multi-layer perceptrons for the grounding of symbols on categories. In addition, it will deal with some problems of current connectionist architectures, such as the scaling up of categories and symbols.