Sinirsel ateşleme verisinden fitzhugh-nagumo noron modelinin parametre kestirimi
Loading...
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Open Access Color
OpenAIRE Downloads
OpenAIRE Views
Abstract
Bu tezde Fitzhugh-Nagumo sinir h¨ucresi modellerinin parametrelerinin sinirsel ates¸leme verisinden kestirilebilmesine y¨onelik bir aras¸tırma yapılmaktadır. S¨oz konusu modelde girdi bir elektrik akımı olup uyaranı temsil etmekte olup c¸ıktı olarak ise ates¸leme hızı modelden alınmaktadır. Konvansiyonel sistem tanılama y¨ontemlerinde kars¸ılas¸ılan durumlardan farklı olarak elde edilen ¨olc¸ ¨umlerde s¨urekli ¨orneklenmis¸ bir veri (zar potansiyeli ya da ates¸leme hızı) s¨oz konusu de˘gildir. Tam tersine, sadece aksiyon potansiyeli zamanlarından olus¸an ayrık bir veri toplanmaktadır. Di˘ger ¨onemli bir ¨ozellik ise bu verilerin iyon kanallarının istatistiksel s¨urec¸leri nedeniyle rastgele olus¸udur. Varıs¸ rastgele s¨urec¸lerinin istatistiksel tanımlanabilirlikleri sayesinde model parametrelerinin kestirimi ic¸in olabilirlik fonksiyonlarının tanımı yapılabilmektedir. Benzetimler sırasında ya sinirsel ates¸leme zamanları modelin c¸ ¨oz¨um¨u yoluyla elde edilmeli ya da bir deneyden gerc¸ekc¸i veri toplanmalıdır. Algoritma sınanması amacıyla birinci y¨ontem tercih edilebilir. Burada parametreleri bilinen modelden elde edilen ates¸leme hızı verisi, homojen olmayan Poisson s¨ureci benzetimi yapılarak sinirsel ates¸leme verisine d¨on¨us¸t¨ur¨ul¨ur. S¨oz konusu benzetimlerde ¨onceden tanımlanmıs¸ bir uyaran profiline gereksinim vardır. Bu c¸alıs¸mada Fourier serisi bic¸iminde tanımlanmıs¸ uyaran profilleri s¨oz konusu olmaktadır. Ayrıca istatistiksel yeterlilik sa˘glanması ic¸in benzetimler c¸ok defa tekrarlanmaktadır. Bu is¸lem sırasında faz ac¸ıları rastgele atanarak benzetimlerin ba˘gımsızlıkları garanti altına alınmıs¸tır. Benzetimlerden elde edilen uyaran/cevap verisi yerel Bernoulli s¨urec¸lerinden t¨uretilmis¸ homojen olmayan Poisson olabilirlik fonksiyonları ¨uzerinden n¨oron parametrelerinin en y¨uksek olabilirlik kestirimi yapılmaktadır. Kestirimi yapılan parametrelerin ortalama de˘gerleri tablolar halinde, istatistiksel ¨ozelliklerinin de˘gis¸imi de grafikler halinde sunulmaktadır. S¨oz konusu grafikler kestirimin standart sapmalarının Fourier serisi uyaranın alt eleman sayısı, taban frekansı, genli˘gi ve ¨ornekleme (tekrarlanmıs¸ benzetim) sayısına kars¸ın de˘gis¸imini incelemektedir. T¨um bunların yanı sıra, gelis¸tirilen yaklas¸ımların performansını inceleyebilmek ic¸in dıs¸ kaynaklardan gerc¸ekc¸i uyaran/cevap verisi (g¨ok sineklerinin H1 g¨orme sisteminden alınmaktadır) alınmıs¸ ve gelis¸tirilen algoritmalar bu verilerle denenmis¸tir. Burada sineklerin g¨orme sistemleri renksiz g¨ur¨ult¨u bic¸iminde uyaranlarla 20 dakika boyunca uyarılmıs¸ ve sinirsel ates¸leme verileri toplanmıs¸tır. Bu deneme aynı zamanda Fourier serisi dıs¸ında bir uyaran ile c¸alıs¸abilme olana˘gı da sunmus¸tur. Bu ac¸ıdan algoritmaların daha genel bir testine de olanak sa˘glamıs¸tır. C¸ alıs¸mada kullanılan hesaplama ortamı MATLAB olup, en iyileme (optimizasyon) k¨ut¨uphanesinde bulunan fmincon beti˘gi olabilirlik kestiriminde kullanılmaktadır.
In this thesis, we attempt to estimate the parameters of a single Fitzhugh-Nagumo neuron based on the neural spiking data. In this model, the input is an electric current serves as the stimulus while the output is considered to be the firing rate of neural spiking. The di erence from the conventional system identification techniques is that no continuous variation of the response (the membrane potential or firing rate) is available. Instead, the data consists of the peak timings of action potentials called as spikes. One major property of these is that they are generated as a result of stochastic processes (ion channel stochasticity). Thanks to the arrival processes in statistics one can implement likelihood functions to estimate those parameters. In the simulation frame work one needs either to simulate the neural spiking or use a set of spike trains obtained from realistic data. For algorithmic testing of the methodologies developed in this research an inhomogeneous Poisson process is simulated using the firing rate response of a Fitzhugh-Nagumo model with known nominal parameters. The firing rate response is obtained from a predefined stimulus which is in Fourier series form with superimposed cosines. The simulations are repeated multiple times with di erent stimulus phases (phases of cosine functions in Fourier series) to obtain enough statistical content. The simulated stimulus-response data is then provided to the inhomogeneous Poisson likelihood functions (derived under Local Bernoulli approxiiii mation) to obtain an estimate of the neuron model parameters. The mean estimated values are presented as tables and their statistical analysis are presented graphically. The graphs present the variations of the standard deviations of the estimates against di erent values of stimulus component sizes, base frequency, amplitude and number of samples. In addition, in order to validate the performance of the methodologies developed in this thesis a realistic stimulus/response data is obtained from external sources (H1 neurons of blowflies) and the algorithms are applied. Here the vision system of the flies are stimulated by a 20 minute white noise stimulus and the neural spikes are collected. It is also convenient to test the algorithm with a di erent set of data other than Fouries series based ones. The computational environment is based on MATLAB and its constrained optimization routine fmincon is used in the likelihood estimation.
In this thesis, we attempt to estimate the parameters of a single Fitzhugh-Nagumo neuron based on the neural spiking data. In this model, the input is an electric current serves as the stimulus while the output is considered to be the firing rate of neural spiking. The di erence from the conventional system identification techniques is that no continuous variation of the response (the membrane potential or firing rate) is available. Instead, the data consists of the peak timings of action potentials called as spikes. One major property of these is that they are generated as a result of stochastic processes (ion channel stochasticity). Thanks to the arrival processes in statistics one can implement likelihood functions to estimate those parameters. In the simulation frame work one needs either to simulate the neural spiking or use a set of spike trains obtained from realistic data. For algorithmic testing of the methodologies developed in this research an inhomogeneous Poisson process is simulated using the firing rate response of a Fitzhugh-Nagumo model with known nominal parameters. The firing rate response is obtained from a predefined stimulus which is in Fourier series form with superimposed cosines. The simulations are repeated multiple times with di erent stimulus phases (phases of cosine functions in Fourier series) to obtain enough statistical content. The simulated stimulus-response data is then provided to the inhomogeneous Poisson likelihood functions (derived under Local Bernoulli approxiiii mation) to obtain an estimate of the neuron model parameters. The mean estimated values are presented as tables and their statistical analysis are presented graphically. The graphs present the variations of the standard deviations of the estimates against di erent values of stimulus component sizes, base frequency, amplitude and number of samples. In addition, in order to validate the performance of the methodologies developed in this thesis a realistic stimulus/response data is obtained from external sources (H1 neurons of blowflies) and the algorithms are applied. Here the vision system of the flies are stimulated by a 20 minute white noise stimulus and the neural spikes are collected. It is also convenient to test the algorithm with a di erent set of data other than Fouries series based ones. The computational environment is based on MATLAB and its constrained optimization routine fmincon is used in the likelihood estimation.
Description
Keywords
Matematik, Mathematics
Turkish CoHE Thesis Center URL
Fields of Science
Citation
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
0
End Page
77