Research On Bio-inspired Models And Algorithms Of Spiking Neural Networks

Despite the widespread applications of artificial intelligence across diverse domains,a significant performance gap between current artificial intelligence and human intelligence still exists.To bridge this gap,brain-like computing endeavors to simulate the structure and functionality of biological brains,thereby developing intelligence systems with brain-level competence.As a pivotal subset of brain-like computing,Spiking neural networks(SNNs)precisely emulate the spike emission mechanism of biological neurons,which has notable advantage in processing temporal data.Nonetheless,many spiking models and algorithms predominantly rely on conventional artificial neural networks,resulting in the underutilization of their temporal processing ability and an evident performance gap compared with human brain.It is of pivotal significance for brain-like computing community to fully exploit the potential of SNNs,enhance their information representation and processing capability,further leverage their temporal processing advantages,combine biological mechanisms and scientific discoveries at the mean time,thereby explore and build more intelligent brain-like models to bridge the gap between artificial intelligence and brain intelligence.To address these issues,with a primary focus on the models and algorithms of spiking neural networks,taking inspirations from biological structures and functions,incorporating brain bionic mechanisms and cross-region information transmission,several bio-inspired models and algorithms are proposed for spiking neural networks.The principal innovations are outlined below:(1)To address the challenge of catastrophic forgetting in traditional SNN learning algorithms,a novel algorithm called the recursive least squares-based learning rule(RLSBLR)is proposed.This method aims to enhance the accuracy,efficiency,and robustness of sequence learning.RLSBLR constructs the error function on the basis of the difference between the membrane potential and firing threshold.Moreover,it incorporates a mechanism to recalculate past errors using current weights,ensuring the preservation of learning integrity.Furthermore,taking inspiration from biological delay transmission mechanisms,delay learning is incorporated into the algorithm to resolves the learning inability within silent windows,thus improving learning performance in sparse spiking sequences.Results from simulation experiments,speech recognition and fault diagnosis experiments have testified the superior learning performance of the RLSBLR algorithm compared to classical sequential learning algorithms.(2)Aiming at both multi-sequence and multi-time step memory,a bionic spiking temporal memory(BSTM)model is proposed with inspiration from the hippocampal structure.This model aims to enhance the accuracy and robustness in sequence memory retrieval.Existing models often produce redundant or erroneous information in multi-sequence prediction.Moreover,these models need to specify the number of predictive steps in multitime step prediction,which restricts their adaptability and practical utility.The BSTM model takes insights from the organization of pyramidal neurons and neural columns in the brain,as well as the hippocampus’ s information transmission mechanisms.By simulating the encoding,storage,and retrieval processes of sequence memory,BSTM has effectively realized multi-sequence and multi-time step memory.Experimental results have demonstrated that the proposed BSTM outperforms other sequence memory models in terms of accuracy and robustness.(3)With inspirations from the neuronal information transmission in the hippocampus and cerebral cortex,on the basis of BSTM with inhibitory and neural oscillation mechanisms incorporated,a new bionic spiking sequence memory(BSSM)model is proposed,which further enhance memory retrieval accuracy and robustness.With competition introduced among neurons and neural columns via inhibitory synapses,the influence of noisy or irrelevant neural activity is reduced,improved feature selection and optimized information transmission.Through the oscillation mechanism,most cortical neurons tend to emit spikes near oscillation peaks,which emulates the oscillation phenomenon in the brain,and enhances neuron synchronization and boosts information transmission efficiency.Experimental results from text retrieval applications demonstrate the superior retrieval accuracy and robustness of the BSSM model over the minicolumn-based models,the SNN-based models,and the traditional deep leaning models.(4)To delve into the processing of motion information from the MT to MST regions in the brain,a novel SNN-based MT-to-MST motion recognition model is proposed.This model is inspired by the information transmission and hierarchical organization from MT to MST region in dorsal pathway brain regions.Combining the electrophysiological data of MT and MST neurons in the monkey brain,the proposed model can effectively extract the information of the form,head orientation,and walking direction of biological motion after learning through synapses.The experimental results show that the MST neurons in the SNN model exhibit different responses to different motor stimuli,and their firing spike patterns and responses closely align with observations from biological studies,affirming the biological fidelity of our model in processing motor information.By analyzing the results of the SNN-based motion model,the hypotheses on the function of neurons in different regions has been guessed,which provides insights and guidance for further biological experiments.(5)Addressing the underutilization of biological mechanisms and specific neuronal functions in conventional visual region models,a retina-inspired deep model(RDM)for image dehazing is proposed.Inspired by the hierarchical structure and neuronal functions of the biological retina,the RDM simulates the information processing from the photoreceptor neurons to the ganglion neurons.It incorporates various biological mechanisms,including lateral inhibition,antagonistic center-surround receptive fields,color-opposition,disinhibition and firing spikes.By combining color and brightness information,this model directly generates clear,haze-free images,effectively reducing the haze.Experiments results on image dehazing datasets D-Hazy and SOTS show that the RDM model achieves comparable performance to classical dehazing models.