Abstract
Computational Learning is an essential part of Intelligent System. There are number of learning approaches and concepts for gathering knowledge, for knowledge management and effective utilization of obtained knowledge. We do believe that this is a crucial part of future autonomous systems. Incremental learning, where number of Agents are gathering, sharing and utilizing information and knowledge seems to be simple but very ambitions solution for Intelligent Systems. Knowledge must be incremental, replicable and easy to read. If learning is the process of acquiring knowledge through study, experience, and teaching, how these three fashions match learning methods studied in the field of computational intelligence? Study -- memorization or reading can be projected into training of Artificial Neural Networks (ANN), which is provided off-line, using training-data file and an error function. Experience is more like a reinforcement learning concept, where knowledge is gained in interaction with environment according to the policy and prescribed value function. Teaching is the most interactive process from these three, it is mutual involvement of learner and teacher, with observation and guidance, without fixed prescribed plan. The closest match might be the Interactive Evolutionary Computation (IEC), which searches for the optimum in interaction with a human according to his or her subjective evaluation of observed results. However, IEC was probably not used in the context of learning systems yet. Ideas behind IEC might be introduced into ANN field to develop interactive learning methods for neural networks. Observation phase from IEC might be implemented using visualization techniques for neural networks, which were already developed but did not gain popularity in application areas. Selection phase from IEC, where candidates for future evaluation are chosen, can be translated into selection of perspective trends in neural network behavior. We visualize responses of particular neurons in network to its inputs, and selectively reinforce these neurons, which are favored by a human observer. This reinforcement of individual neurons can be realized by several means, amplification of outputs of neurons, bias shift, modifications of learning rate, or even reinitialization of weights. These techniques have to be further studied in order to find optimal mixture of them. We also study clustering of neurons into groups, which is necessary for visualization of practically big enough networks. Hopefully, interactive learning methods will help in the great task of computerized search of global minima, but also to bring artificial neural networks into new application areas, these which are more subjective, and more human.
