Spike-based Coding And Learning Algorithm For Spiking Neural Network And Its Applied Research

Biological evidence suggests neurons in the nervous systems transmit information through action potentials(or called as spikes).Thus,the third generation of neural network,spiking neural network(SNN)has gained a lot of attention.Unlike the traditional frequency-based neural networks,SNN is able to process and extract features from the temporal dynamics encoded in the spike signals,thus making it more biologically plausible,and has greater computational power and lower power consumption,and has broad application prospects in the future mobile intelligence field.The fundamental computation of single neurons is the transformation of incoming spike trains into appropriate spike output.However,it is still mysterious that how neurons with spiking features give rise to powerful cognitive functions of the brain.The current achieve-ments have initially demonstrated the powerful functions of SNN,but the research on SNN is still in its infancy.It lacks effective neural information coding methods,and the effec-tiveness and applicability of existing learning algorithms cannot be guaranteed due to the complexity in encoding and the undifferentiability of the spiking variables.This thesis presents detailed investigation on information processing and cognitive com-puting in SNN,trying to reveal and utilize mechanisms how the biological systems might operate.Temporal coding and learning are two major concerns in SNN,with coding de-scribing how information is carried by spikes and with learning presenting how neurons learn the spike patterns.The focus of this thesis varies from a neuronal level to a system level,including topics of spike-based sensory coding and learning in single and multilayer neural networks,system modeling,as well as applied development of visual and auditory processing systems.The main contents and innovations of this paper are as follows.1.This paper proposes spike-based encoding and learning of spectrum fea-tures for robust sound recognition,which treats the auditory coding and learning as a systematic process.This method not only improves the recognizability of sound in the case of noise,but also better simulates the process of sound sig-nal processing and learning in the auditory cortex of the nervous system.Most of conventional methods use stationary(frequency-based)features which are not robust to noise,as each stationary feature contains a mixture of spectral information from both noise and signal.This paper proposes an spike-timing based integrated model for robust sound recognition,which is a step forward towards the design of encoding and learning the local time-frequency(LTF)features extracted from sound spectrogram using SNN.In this model,we select the local maximum value(high energy peak)as keypoints(which contain the LTF information),and encode the information into spiking train.We further analyze the efficacy of the spike-based feature encoding method and the recognition performance of the model by using two classes of SNN learning algorithms,respectively.Utilizing the temporal coding and learning,networks of spiking neurons can effectively perform robust sound recognition tasks.Experimental results demonstrate that the model achieves superior performance in mismatched conditions compared with benchmark approaches.2.A unified and consistent spike event-driven categorization model for Address-Event Representation(AER)image sensors is proposed,which could encode and learn the temporal information based on fully precise-timing spikes.Through the network,we fill the gap between a real-world problem(image en-coding)and diffcerent learning algorithms for AER image sensors.Most existing architectures for object recognition,spike event-based feature extraction and learning are studied separately,while this model is important in the light of recent trends in combining both the coding and learning in a systematic level to perform cognitive computations.To solve this problem,we present a systematic computational model to explore brain-based computation for object recognition.This model is consistently implemented in a temporal framework,where the fully precise-timing spikes(multi-spike)processing is considered for temporal feature encoding and learning;Otherwise,a noise reduction method is also pro-posed by calculating the correlation of an event with the surrounding spatial neighborhood based on the recently proposed time-surface technique.The model evaluated on a wide spec-trum datasets demonstrates its superior recognition performance especially for the events with noise.Such a model would be beneficial for applications in both hardware and software implementations.3.A supervised learning algorithm for learning precise timing of multi-spike in multilayer spiking neural networks is proposed,which can efficiently and robustly process the complex spatio-temporal spike patterns.The majority of research about multispike learning methods has focused on training single rather than multilayer networks,since the complexity of the learning targets increases significantly for multispike learning.It is still a core challenge to trigger multiple precise timing spikes in each layer of multilayer spiking neural network?The computational efficiency of existing algorithms is low.To address this issue,we propose a novel supervised,multispike learn-ing method for multilayer SNN,which can accomplish the complex spatio-temporal pattern learning of spike trains.The proposed method derives the synaptic weight update rule from the Widrow-Hoff(WH)rule,and then credits the network error simultaneously to preceding layers using backpropagation.Especially,both the time-driven and event-driven compu-tation mechanisms are used for simulating neural models in this work.The algorithm is successfully applied to the XOR problem and the UCI datasets,as well as to complex noise problems.Experimental results show that the proposed algorithm can achieve comparable classification accuracy with classical learning methods and a state-of-the-art supervised al-gorithm.In addition,the training framework effectively reduces the number of connections,thus improves the computational efficiency of the network.4.This paper proposes a simple and novel threshold-driven multi-spike learning algorithm for spiking neurons,which can quickly and accurately per-form various classification and temporal credit assignment(TCA)tasks.A multi-spike output learning rule can train neurons to fire desired number of spikes,which empowers them to discover sensory features embedded in a complex background activity.However,the formulation of efficient supervised learning algorithms for spiking neurons is complicated and remains challenging.Most existing learning methods with the precisely firing times of spikes often result in relatively low efficiency and poor robustness to noise.To address these limitations,we propose a simple and effective multi-spike learning rule to train neurons to match their output spike number with a desired one.The proposed method will quickly find a local maximum value(directly related to the embedded feature)as the relevant signal for synaptic updates based on membrane potential trace of a neuron,and constructs an error function defined as the difference between the local maximum membrane potential and the firing threshold.With the presented rule,a single neuron can be trained to learn multi-category tasks,and can successfully mitigate the impact of the input noise and discover embedded features.Experimental results show the proposed algorithm has higher precision,lower computation cost,and better noise robustness than current state-of-the-art learning methods under a wide range of learning tasks.