A Biologically Inspired Method for Conceptual Imitation Using Reinforcement Learning

Levels Observations are categorized to concepts with respect to some principles that depend on physical and=or functional characteristics of the items. From this perspective, Zentall et al. have categorized concepts to three levels of abstraction (Zentall et al. 2002) (see Figure 1): Perceptual: These concepts are formed solely by measuring similarity of instances in perceptual space. Such data can be categorized by simple clustering algorithms in an unsupervised fashion. Relational: In this type of concept, although perceptual similarity still contributes to categorization, it is not sufficient to form the correct concepts. External information must link perceptual categorizes and form the right concept. This is achieved by classical conditioning. Associative: In learning these concepts, the stimuli within classes bear no obvious physical similarity to one another, but cohere because of shared functional properties. Same/Different Judgment As we move from perceptual toward associative concepts, more complex cognitive capabilities are required. Fortunately relational concepts FIGURE 1 Three types of concept: perceptual (left), relational (middle), and associative (right) in a two-dimensional feature plane. Biologically Inspired Reinforcement Learning 159