Self-organizing in Neural Networks Based on Memorizing

The goal of artificial intelligence is to develop a system that could deal with intelligent tasks in autonomous way. This goal is not achieved, but through this research many interesting algorithms have been investigated and developed, and each of them can provide some specific aspect of intelligence. For example, theorem proving [1], task solving [1], expert systems [1], pattern and speech recognition [1], clustering [1], decision tree learning [2], adaptive behavior in reinforcement learning [2], deterministic automata learning [3], error minimization in neural networks [4], and so on. However, if we would like to build an autonomous adaptive intelligent system (for example, adaptive autonomous robot), we have to combine many of developed approaches into a single learning algorithm. And usually we are not able to perform this combination in natural way, so a very sophisticated system is obtained, that could not be robust. Artificial intelligence research has been inspired by biology because intelligence phenomena were mostly observed in human beings. However, all these approaches have a weak relation with biological systems. What does it mean? It means a potential danger – one day the selected direction of developing alternative, to natural, intelligence can find itself in impasse. One of the most biologically based approaches in artificial intelligence is Kohonen neural network [4], however it omits implementation of many important phenomena, for example, emotional control of neural network. Our investigation is connected with a new artificial neural network, proposed by Emelyanov-Yaroslavsky in 1990 [5]. We will refer this neural network by a term “inductive automaton”. The main idea was not to implement a specific functionality for each class of tasks but to define one basic task – energy consumption minimization and then to obtain all other functionalities as consequences of it, as by-products. The second criterion of developing that network was looking for a strong analogy between the artificial network and natural neural systems. Therefore, it is a good example of trying to solve two previously discussed problems.