Marco Mazzone & Alessio Plebe - Differential emergence of two concept types in an artificial neurocortical model.

Although there are good reasons to attribute concepts also to animal species very different from our own, it is far from clear whether the conceptual abilities of other animals are comparable, in quantity and in quality, to those of the human species. Moreover, the complex hierarchy in types of conceptual structures in humans is quite probably unique. In fact, humans alone come to recognize a huge variety of objects, properties and relations. On the other hand, humans alone make use of sounds to identify and categorize objects, properties and relations. This correlation could be principled. Interestingly, some authors have reported that the ability to use the non-obvious cue of naming in order to categorize objects emerges at about the same age as the so-called ‘vocabulary spurt’ (18 months). One might speculate that this ability could be a precondition for the capacity to form a rich conceptual system, and therefore, that the vocabulary spurt is indeed intimately connected to this sort of ‘conceptual spurt’. In particular, we suggest that language plays a crucial role in the emergence of concepts concerning the properties of objects. A large literature on neurocomputational modelling has addressed the issue of the emergence of categorization from statistical regularities in the inputs (McClelland & Rogers, Semantic Cognition: A Parallel Distributed Processing Approach, Bradford Books, 2006); there are also models of the emergence of the correlation between words and concepts. In general there are, however, two orders of weakness in those models. First, they lack a plausible correspondence with biological computational structures. Second, they are generally vague in accounting for the emergence of different types of concepts. Here we adopt a neural network model which attempts to simulate, in a biologically plausible way, the process by which infants learn how to recognize objects and words through exposure to visual stimuli and vocal sounds. The model is a hierarchy of artificial cortical maps that are analogous to the brain circuitry composed by the visual ventral stream, the auditory path and areas in the prefrontal cortex. The artificial maps evolve specific functions as a result of their plastic Development driven by environmental visual and vocal stimuli, as well as intrinsic activity. This group of functions demonstrate to support the emergence of levels of concepts types at the higher end. In particular, we use the model to investigate how concepts of color emerge from environmental regularities, including the exposure to color names. Young children experience considerable difficulty in learning color terms, one of the first concept types different from referential names. The model is first trained on correlations between a set of 100 objects and 38 category names (our complete training set of vocal stimuli includes the 7200 most common English words). Subsequently, correlations between objects and color names are introduced in a variety of conditions. This procedure appears quite promising in order to test debated issues in the literature on concept/meaning formation, such as the question of how the infant’s preference for whole objects can be overcome.