Development of new learning algorithms for spiking neural networks

The main shortcomings with the gradient descent-based learning algorithms used for neural networks are that the convergence speed is relatively slow and the algorithm can be easily trapped into a local optimum. The studies that demonstrate the robustness of variable structure control have motivated the use of the sliding mode control approach in the training of Arti cial Neural Networks (ANNs). In this dissertation, the development of a Spiking Neural Network (SNN) with a novel variable structure systems (VSS)-based learning algorithm is considered for the identi cation and control of dynamic plants. The parameter update rules are derived based on the Lyapunov stability method. Stable online tuning of the neurocontroller parameters and a fast learning speed are the prominent characteristics of the proposed algorithm. Neural networks are very e ective in implementing a mapping between the inputs and outputs of a system, but they have a black box nature. On the other hand, fuzzy logic systems are good at explaining how they reach their decisions but they require expert knowledge. If the knowledge is incomplete, wrong or contradictory, then the fuzzy system must be tuned. In the proposed fuzzy SNN (FSNN) structure, SNN is utilized to automate this process and substantially reduce development time and cost while improving performance. The update rules have been derived based on the VSS theory to avoid global stability problems. The developed algorithm has been successfully implemented for the wheel slip regulation of an Antilock Braking System (ABS) and trajectory control of a two-dof SCARA manipulator.