Article Preview
TopIntroduction
The second part of the 20th century can be considered as the ‘golden age’ of cognitive psychology and cognitive neuroscience; cognitive psychology is based on the founding metaphor of the mind as an information processing device while in cognitive neuroscience the basic assumption is that the information theory approach is crucial in order to derive well founded structure to function inferences connected to brain/neural activity. The mind/brain as an information device metaphor provided a common conceptual framework which allowed productive exchanges between many different disciplines such as psychology and neuropsychology, neurophysiology, artificial intelligence and I.T., theoretical physics, linguistic and neuro-linguistic, philosophy of mind and to some extent, social sciences. An paradigmatic example of the success of this conciliency strategy can be seen in the domain of cognitive informatics (Wang, 2003, 2009a, b; Wang et al., 2011) where information is considered as the crucial variable to be considered in order to model natural/mental phenomena.
Given the peculiar nature of this huge disciplinary convergence the conceptual and empirical tools that were developed in this domain are heterogeneous, including behavioral data, single/multiple cells recordings, human and animal lesions data, imaging evidences, simulations and analytical studies in the domain of both hard and soft computation, analytic modelling based on statistical mechanics theory and linguistic analysis tools. Given this very large base of evidence, the problem of deriving appropriate methods to obtain valuable scientific inference is one of the main issues in cognitive psychology and neuroscience (Shallice, 1988). The development of box-and-arrows diagrams and the cognitive architectures approach is an example of an efficient conceptual device in this respect. Anderson & Lebiere, 2003, elaborating on Newell’s positions, derived 12 criteria that a human cognitive architecture should fit in order to be functional: flexible behavior, real-time performance, adaptive behavior, ability to resort to a vast knowledge base, ability to face a dynamic environment, ability to integrate in an efficient way different sources of knowledge in order to implement high level knowledge manipulation, such as required in order to support inference, induction, metaphor, and analogy, use of ‘natural’ language, learning from the environment, sensitivity to developmental constraints, capacity to exhibit consciousness and self-consciousness and brain realization. They then compared two of the most prominent competitors, ACT-R (Anderson & Lebiere 1998) and Connectionist Models (McClelland, Rumelhart & PDP Research Group, 1986) along these criteria. Here I am not interested in reporting the results of this comparison, as I will focus instead on some of the areas in which both types of models seem to be defective.