Research Of Temporal Sequence Information Recognition Based On Neural Network

Neural network has always been a hot topic of academic research.With the replacement of graphics hardware,the deep learning based on the neural network again has achieved fruitful results in various fields.However,these artificial neural networks don't fully account for the operating mechanism of biological neurons while dealing with information.The Spiking Neural Network,which is the latest research in the field of neuroscience,has a high degree of bio-simulation.Spiking Neural Network can deal with the characteristic information in the spatial-temporal domain well,and encode the external stimulus in the form of temporal characteristic.Finally,the encoded pulse information can be processed by the nervous system.Spiking Neural Network has achieved a lot of practical applications in biological identity authentication,speech recognition and other fields.The brain neurons can recognize the sequences of external stimuli information with spiking features,but its biophysical mechanisms are poorly understood.It is very important to study the temporal information recognition based on spiking neural network,which can help understanding the principle of information processing in the brain and be applied to the recognize and process external complex spatial-temporal information,so the research of this thesis is very promising.In general,the content of this thesis as follows:1.Briefly introduce the theoretical knowledge about the recogniton of temporal inforamtion based on Spiking Neural Network,including the related biological characteristics and Spiking Neural Network model of neuron.Then,study the theory which related to the information coding,training neurons and the construction of neuron model to decoding temporal sequence.2.Propose a new supervised learning algorithm which named DL-PSD.Based on the classical algorithm PSD and the time delay characteristic of the neuron,the DL-PSD algorithm is proposed for the temporal sequence information recognition in the Spiking Neural Network.And DL-PSD effectively improves the efficiency of item recognition of temporal information.3.Propose a new mechanism to decode the Spiking Neural Network sequence and construct the corresponding neuron model.The traditional convolution methoddoesn't make full use of the biological characteristics of neurons while dealing with the recognition of temporal information.Taking into account the basic principles of FSA recognition of handwritten words,We have built a decoding unit model of recognition a specific temporal sequence which combined with the biological neurons with the dendrites bistable plateau potentials.4.Finally,combine the information coding,item recognition and temporal sequence information decoding together to recognize the temporal information.phase coding method is used to transform the image information.Then,we use the DL-PSD to train learning neurons to recognize item and utilize the new decoding mechanism to recognize specific character image sequences.Experiments have successfully identified the specific optical character image sequence,and we can identify many kinds of specific character sequences when changed the connection structure of the learning output neuron and the decoding model sensing unit.And the above verifies the extensibility and robustness of the our recognition mechanism.This provides a new way to construct a common identification structure to encode and process human biometric information,as well as in the field of image processing.