Research On A Kind Of Spiking Neural Network With Hybrid Neuroplasticity

Spiking neural networks(SNNs), referred as the 3rd generation of artificial neural networks(ANNs) increase the level of realism in neural simulations by incorporating the spike mechanism and kinetics of synapses. In early days, the learning mechanisms in SNNs are based on numerical optimization similar to the traditional ANNs. Later, some methods which combined numerical optimization and neuroplasticity learning emerge. In the 1990 s, with the discovery of STDP(spike-timing-dependent plasticity) mechanism of synapses in biological neural systems, neuroplasticity becomes a hot topic in neuroscience.Neuroplasticity is an umbrella term that encompasses both synaptic plasticity and non-synaptic plasticity. The former deals directly with the strength of the connection between two neurons, and affects the network topology. The latter involves modification of neuronal excitability in the axon, dendrites, and soma of an individual neuron, remote from the synapse, but indirectly affects the topology of the network.The optimization of SNN topology should be a combination of several types of plasticity. In this thesis, both synaptic and non-synaptic plasticity are taken into account. The main research focus on the short-term synaptic plasticity, STDP(including excitatory STDP and inhibitory STDP), a non-synaptic plasticity called intrinsic plasticity and their effects on network structure and computing properties.Firstly, by analyzing the combination method of multiple plasticity mechanisms, this thesis proposed a hybrid plasticity SNN(HP-SNN) model. Three indicators in complex network theory: degree distribution, shortest path length, clustering coefficient; synchronization degree of spiking activities; and information entropy are used to evaluate the optimized SNN.To further verify the application performance of the HP-SNN proposed. We apply it to the reservoir in "reservoir computing". Two benchmark tasks are considered: bionic signal reconstruction and classification of jittered spike trains. Experiments results based on MATLAB show that, compared with conventional SNN, HP-SNN reservoir has smaller errors and higher classification accuracy. For the same task, HP-SNN consisting of smaller neurons can achieve the performance of conventional SNN.