Research On Training Algorithm Of Deep Spiking Neural Network Based On STDP

Spiking Neural Network SNN is also known as the third generation of neural network.It realizes synaptic computation by means of pulse event driving,which is expected to break through the bottleneck of energy and throughput faced by artificial neural network at the present stage,and has great potential application value in the fields of robot,autopilot,military,aerospace,etc.But SNN still has problems that cannot be trained directly.The SNN training algorithm based on the Spike-Timing Dependent Plasticity(STDP)rule is one of the main training algorithms for SNN,because of its physiological basis and simple calculation.But there are still some problems on complex datasets,such as low classification accuracy,large inference delay,and difficult to implement deeper layer network.In order to improve the robustness,learning efficiency and performance of SNN training algorithm based on STDP rules,this thesis studies the R-STDP algorithm with reward and punishment signal and the hybrid algorithm with unsupervised and supervised STDP.The main content of this thesis are as follows:(1)An MR-STDP(Mix-Rewarding STDP)algorithm with mixed rewards and punishments is proposed.Aiming at the problem that the feedback signal of R-STDP algorithm only acts on the bottom layer of the R-STDP network,a new MR-STDP model with mixed signal of reward and punishment is formed by using the self-encoder and its unsupervised characteristic.Experimental results on MNIST and COVID-19 CT datasets show that the proposed method achieves higher accuracy than R-STDP,and the learning efficiency of shallow layer network is greatly improved.(2)A hybrid training algorithm combining unsupervised STDP and supervised STDP is proposed.By analysing the non-differentiable terms in back-propagation,the biological meaning of the non-differentiable terms is defined,and the approximate substitutes of the non-differentiable terms with lower computational complexity are proposed.Based on this,an approximate gradient descent training algorithm with STDP rules is proposed and combined with unsupervised STDP to form a hybrid training algorithm with unsupervised STDP.Experimental results show that the proposed hybrid training algorithm not only improves the precision of the impulsive neural network,but also enhances the anti-noise performance of the network.