| Summary: | Artificial intelligence (AI) has been widely used in versatile applications (robot, autonomous vehicle, gaming, industry IoT, etc.), and changed the way that human being lives. Meanwhile, huge amount of data has been created in this big data era in recent years. Neural Network, as an AI technology benefiting from such huge amount of data, becomes mature for commercialization. Successful examples include convolutional neural network (CNN) for image processing and recurrent neural network (RNN) for speech recognition. Compared with the former two technologies, spiking neural network (SNN), as the third-generation neural network, still has a distance from commercial application.
Spiking Neural Network is a neural network which is most similar to biological neural network. The development of neuroscience makes the simulation of the biological neuron model more feasible. From Hodgkin-Huxley model of the early time to the newly raised Spike-Timing Dependent Plasticity (STDP) model, the electrochemical reaction of biological brain has been described more accurately.
In this dissertation, an algorithm of SNN is implemented. Based on the STDP model and some modern dynamic biological neural system model (Homoeostasis, Lateral Inhibition, etc.), this configurable algorithm is used to achieve hand-written digits recognition based on the benchmark of MNIST dataset. A supervised layer is added following the spiking neural layer for better classification. To evaluate the algorithm, experiments of different number of digits classification are carried out and reported.
|
|---|