Concept learning from visual experiences using unsupervised neural networks

Ability of learning and understanding are essential for intelligent systems. Concept formation is a way of representing knowledge in machines as humans do. To achieve true intelligence in machines there is a necessity of developing techniques which can accumulate knowledge from a given sequence of input patterns without human intervention. Feigenbaum's EPAM, Fisher's COBWEB and Lebowitz's UNIMEM are some of the existing models for concept formation. These techniques use a decision tree for concept formation which perform a number of tests for each input. Therefore they have several shortcomings such as instance arrival order, processing time and ad hoc nature of selecting attributes for tests. This paper describe a technique which extracts several common features of the existing concept formation techniques and demonstrates a novel method of extracting concepts from visual experiences using an unsupervised neural network which addresses the above mentioned shortcomings. An unsupervised neural network called Growing Self Organizing Map is used to obtain the clusters from the data set and then a statistical analysis is carried out for extracting the contributing attributes for concepts.