Dekhtyarenko O. Development and analysis of methods for construction of sparse associative neural networks

In this thesis we study the models of associative memory based on sparsely connected Hopfield-type network. The main consideration is given to learning rules that are aiming the maximization of associative properties subject to certain architectural constraints. An improvement of pseudoinverse learning rule is introduced and investigated for the networks with predefined architectures. This improvement eliminated the numerical instability of weights calculation and increased memory capacity of the network. Theoretical estimations of the network associative behaviour and weight matrix properties are obtained. A novel model of network with adaptive architecture that depends on the stored data is proposed. A phenomenon of phase transition into the state of associative recall is discovered and investigated for this model. A new deterministic way of construction of the small-world architecture for associative networks is proposed. It provides better associative properties of the network while preserving allknown advantages of the small-world model. Open source software library implementing associative models and algorithms is created. The efficiency of proposed models of sparse associative networks has been demonstrated in the Electronic Nose application.