Estudo de uma classe de memórias associativas hierárquicas baseadas em acoplamento de redes neurais artificiais

Abstract

Understanding human cognition has proved to be extremely complex. Despite this complexity many approaches have emerged in the artificial intelligence area in an attempt to explain the cognitive process aiming to develop mechanisms of software and hardware that could present intelligent behaviour. One of the proposed approaches is named embodied embedded cognition which through its theoretical-conceptual basis on the cognitive process has contributed, in an expressive way, to the development of intelligent systems. One of the most important aspects of human cognition is the memory, for it enables us to make correlations of our life experiences. Moreover, more recently, the memory process has been acknowledged as being a multi-level or hierarchical process. One of the theories that concerns this concept is the theory of neuronal group selection (TNGS). The TNGS is based on studies on neuroscience, which have revealed by means of experimental evidences that certain areas of the brain (i.e. the cerebral cortex) can be described as being organised functionally in hierarchical levels, where higher functional levels coordinate and correlate sets of functions in the lower levels. The most basic units in the cortical area of the brain are formed during epigenesis and are called neuronal groups, defined as a set of localised tightly coupled neurons constituting what we call our first-level blocks of memories. On the other hand, the higher levels are formed during our lives, or ontogeny, through selective strengthening or weakening of the neural connections amongst the neuronal groups. To account for this effect, we propose that the higher level hierarchies emerge from a learning mechanism as correlations of lower level memories. In this sense our objective is to contribute to the analysis, design and development of the hierarchically coupled associative memories and to study the implications that such systems have in the construction of intelligent systems in the embodied embedded cognition paradigm. Thus, initially a detailed study of the neurodynamical artificial network was performed and the GBSB (Generalized-Brain-State-in-a-Box) neural network model was chosen to function as the first-level memories of the proposed model. The dynamics and synthesis of the single network were developed and several techniques of coupling were investigated. The methods studied to built the second-level memories were: the Hebbian learning, along with it a synthesis based on vector space structure as well as the evolutionary computation approach was employed. As a further development, a more in depth analysis of the storage capacity and retrieval performance considering single networks and the whole system was carried out.