| Abstract [eng] |
Artificial neural networks developed in the scientific field of machine learning are used in practical applications, such as data recognition, prediction of processes and etc. These networks were based on the knowledge of biological neurons which existed at the time. Because of recent advances in neuroscience, it is reasonable to expect the development of a new generation of more effective artificial neural networks in the field of machine learning. The focus of this dissertation is possible applications of spike-timing-dependent plasticity (STDP) for recognition of spike patterns. Models of neuron and STDP were taken from the field of computational neuroscience. A few problems related to this kind of learning were identified and solved during the research. This dissertation proposes three novel models for spiking networks: a neural network which can learn long-lasting sequences of temporal codes; a neural network which is capable of differentiating overlapping spatial patterns and a neural network which can adapt to conditions of variable noise. Also, three different models of STDP spike interactions were compared: nearest neighbor interaction, all-to-all interaction and triplet interaction. It was discovered that under certain conditions the model of triplet interaction is more effective by far than the other two. |
