Modelling Cell Assemblies

Neural Networks are very popular computational models that are generally said to be inspired by human neural functioning. However, neurons in most neural nets function quite diierently from natural systems. This paper describes a novel architecture, the CANT (Connections, Associations and Network Technology) model which is designed to function like natural neural systems. It rst elucidates the important aspects of the model and how they relate to natural systems. The basic idea is derived from D. O. Hebb's idea of the Cell Assembly which is the neural equivalent of a concept. The paper goes on to describe three instances of the model. In each case, stimulus is presented to the network by directly activating neurons in the system. In any given run of the network, this should lead to activation of the appropriate CA. When learning is involved, early runs on the system will not lead to activation of a CA. Instead, the early activation will lead to the system learning the CA. CAs are initially determined by appropriate levels of activation in the network. Later they are determined by statistical analysis of the activation patterns. The rst experiment shows that the CANT model generates CAs given a wide variety of parameter settings. Learning is essential in neural models; the second experiment shows how learning can develop a CA where mere parameter setting can not. In natural neural systems, multiple CAs must exist as natural systems must have multiple concepts. The third experiment shows how a simple net can contain two CAs. The paper concludes with a discussion of future developments of the model. These include duplicating psychological and neurobiological data. This will require further analysis of CAs, learning algorithms and development of higher order formalisms.