Research On A Bionic Model Of Olfactory Systems And Its Applications

It is an important direction of intelligence research area to construct bionic models sim-ulating biologic neural systems.The dissertation studies a set of novel olfactory neural networks,K-set models proposed by Prof.Walter J.Freeman,and applies the Kâ…¢model on some engineering applications.K-set models include K0,Kâ… ,Kâ…¡and Kâ…¢models, among which Kâ…¢model factually simulates the whole olfactory neural network.Firstly,this dissertation analyzes the Kâ…¢model structure using the small-world network theory.Based on wavelet transform,the changes of clustering coefficient and characteristic path length with the edges rewired probability are analyzed.Analytic results show that the Kâ…¢model structure has the small-world network characteristic partly.Then,pattern recognition based on stand-alone Kâ…¢model is studied.Based on phase diagram,analytic results show that Kâ…¢model can form unique landscape of limit cycles corresponding to different patterns no matter the input is in a certain level or variable.For level inputs,stand-alone Kâ…¢model is applied to three typical pattern recognition applications(face recognition,text classification and speech recognition), where feature extraction is needed,based on conventional pattern recognition process. For face recognition,a new face feature extraction method is extracted based on sub-image feature combination idea.According to experimental results,it is shown that Kâ…¢model has powerful pattern recognition capability and can recognize more 40 pat-terns.Otherwise,for time-varying signals with variable input,Kâ…¢model can recognize time-varying signals directly through suitable segmenting methods and the feature ex-traction is not necessary.According to the predefined metric,the Kâ…¢model is applied to recognize two kinds of time-varying signals,EEG and speech.Experimental results show that Kâ…¢model can recognize different time-varying signals effectively.Accord-ing to these different applications,it can be concluded that Kâ…¢model has the potential to be an universal classifier.Thirdly,the dissertation studies the Kâ…¢model combining with transductive confi-dence machine(TCM)and applies it to two applications,EEG and odor(E-nose signal recognition)classification.Compared with stand-alone Kâ…¢model,experimental results show that Kâ…¢model combining with TCM can keep good accuracy and synchronously gives the confidence value of prediction.Finally,a simplified bionic model of olfactory systems is proposed to generate texture images.Compared with the feed-forward net-work based on back-propagation algorithm,the proposed model has higher efficiency to generate texture images and the generated texture images are more abundant and beautiful.Experimental results also show that introduced noise can improve the qual-ity of generated texture images.