guest ::
| Home > Publications database > Statistical Evaluation of Dithering Methods for Pattern Detection |
| Home > Publications database > Statistical Evaluation of Dithering Methods for Pattern Detection |
| Master Thesis | FZJ-2021-00161 |
; ;
2020
RWTH Aachen
Aachen
Please use a persistent id in citations: http://hdl.handle.net/2128/26708
Abstract: The study of neuronal activity patterns is one of the high-interest topics in neurosci-entific research (Abeles (2010)). It is hypothesized that highly interconnected groupsof neurons, so-called cell assemblies, can be building blocks for information processing(Hebb (1949); Abeles (1991); Harris (2005)). One of the effects of these cell assemblies’activation would be the formation of precisely timed arrangements of spikes emitted bythe neurons participating in an assembly (Bienenstock (1995); Izhikevich (2006)). Ex-perimental evidence for millisecond precise spiking activity is available for synchronypatterns (Riehle et al. (1997); Kilavik et al. (2009); Torre et al. (2016)) and for patternswith temporal delays (Prut et al. (1998); Villa and Abeles (1990); Russo and Durstewitz(2017)).One of the algorithms developed in recent years to verify this hypothesis is calledSPADE (Torre et al. (2013, 2016); Quaglio et al. (2017); Quaglio (2019); Stella et al.(2019)), which stands for Spike PAttern Detection and Evaluation. This method com-pares electrophysiological data against a null hypothesis. Due to the complexity of thespiking recordings, this null hypothesis is not determined analytically but obtained by aMonte Carlo approach, i.e., by opportunely generated surrogates (Torre et al. (2013)).This thesis aims to compare different surrogate methods, specifically regarding theirapplication to form a statistically robust null hypothesis. Thereby, we refer to a sub-category of surrogate approaches called dithering methods (Louis et al. (2010b)), whoseparticular feature is that each spike is displaced individually within a small window.We proceed in such a way that first, we summarize the necessary neuroscientificbackground in chapter 1. After a brief explanation of the composition of neurons, wewill introduce action potentials, or spikes, that serve to propagate information. Wecontrast the two standard hypotheses for neural coding and provide the reasons behindthe expectation to find spatio-temporal patterns.The second chapter gives an overview of the theory of statistical point processesused to model spike trains, i.e., sequences of spike times. We introduce the commonstatistical measures used to describe electrophysiological recordings. In particular, wefocus on renewal and Markov processes, which allow simple modeling of stationary andnon-stationary spike trains. Throughout the thesis, we consider mainly the Poissonprocesses with refractoriness (PPR), an adaptation of the Poisson processes, and theGamma processes. In detail, we discuss their statistical properties. Furthermore, wepresent a typical form of discretization of time (used in SPADE) and point out the3significant implications of this approach.We describe in the third chapter the particular complexity of spike trains obtainedin electrophysiological recordings. As an example, we refer to the reach-to-grasp dataset (Brochier et al. (2018)), which was already used for analyses with earlier versions ofSPADE (Torre et al. (2016)).SPADE is presented explicitly in the fourth chapter. We do not only provide anoverview of the entire workflow and how spatio-temporal patterns are defined, but alsoexplain in detail how the patterns are mined and statistically evaluated.Chapter 5 presents the three dithering methods and compares them against one an-other. We contrast the standard approach of uniform dithering with a newly introducedadaptation, which takes refractoriness into account, and with the joint-ISI dithering. Foruniform dithering, we analytically examine the generated surrogate spike trains in termsof their statistical properties. The same statistical properties are determined numeri-cally for the other two methods, enabling us to compare the performance of the threemethods.In the last chapter, we develop a test case that allows a final assessment of theperformances of the presented surrogate methods within SPADE, by evaluating them interms of false positives and false negatives. We can compare the results obtained by thesurrogate methods against a ground truth since we generate spike trains by employingthe processes presented in the second chapter. Therefore, we discuss the reasons forthe occurrences of false positives and false negatives. Consequently, depending on thefiring rate, we recommend which dithering method to use for statistically robust patterndetection.In conclusion, we summarize the results of this thesis and place them in the broadercontext of other surrogate methods, including non-dithering methods. Finally, we ad-dress the research questions raised by this thesis, and we give an outlook concerningupcoming studies.
|
The record appears in these collections: |
PDF
Fulltext by OpenAccess repository