Kolektivna dinamika fizioloških signala

master's thesis

• Title (english): Collective dinamics of physiological signals
• Author: Domjan Barić
• Mentor: Davor Horvatić (mentor)
• Committee member: Davor Horvatić (predsjednik povjerenstva)
• Committee member: Krešimir Kumerički (član povjerenstva)
• Committee member: Nikola Poljak (član povjerenstva)
• Committee member: Željko Skoko (član povjerenstva)
• Committee member: Danko Radić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Physics); Zagreb
• Defense date and country: 2018-12-18, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Physics

Abstract

U ovom radu cilj je analizirati smo podatke dobivene snimanjem aktivnosti neurona transgeničnog miša (ekspresija GCaMP3, genetski enkodiran kalcij indikator) tijekom trčanja po pokretnoj traci. Koristeći Isingov model za analizu dinamike spinskog lanca, želimo pronaći vjerojatnost svakog stanja i razumjeti strukturu mreže neurona. Prvi korak je riješiti inverzni problem te pronaći parametre Isingovog modela. Osim egzaktne optimizacije parametara Isingovog modela, primjenjujemo nekoliko aproksimativnih metoda koje uspoređujemo s egzaktnom optimizacijom. U drugom dijelu rada opisujemo kako su metode strojnog učenja prirodno produženje inverznih problema u fizici. Koristeći strojno učenje možemo definirati metriku točnosti. Možemo predvidjeti u kojem će se slijedećem stanju neuron naći s točnošću od 92%. Opisan je algoritam koji predviđa stanje cijele mreže neurona, a ne samo jednog neurona. Za mrežu koja se sastoji od 37 neurona, možemo predvidjeti slijedeće stanje mreže s greškom manjom od četvrtine neurona po primjeru. Pomoću metoda strojnog učenja klasificiramo svako pojedino izmjereno stanje neurona u stanje rješavanja zadatka i stanje nagrade. Osim što s točnošću većom od 95% možemo klasificirati stanja, možemo odrediti aktivnost kojih neurona signalizira da se radi o izvršavanju nagrade ili zadatka.

Abstract (english)

In this thesis, we analyzed neural activity data obtained from transgenic mice (expression GCaMP3, genetically encoded calcium indicator) during tasks/reward sessions. Using Ising’s model for analysis of spin chain dynamics, we want to find the probability of each state and understand the structure of the network. The first step is to solve the inverse problem and find the parameters of the Ising model. In addition to the exact optimization of the Ising model parameters, we apply several approximate methods that we compare with exact optimization. In the second part, we describe how machine learning methods are a natural extension of inverse problems in physics. Using machine learning, we are able to not only answer various questions, but we can define a metric of precision. To the question in which state will neuron be in the next moment, we are able to answer with the precision of 92 %. We also describe an algorithm that predicts the state of the entire neuronal network, not just one neuron. For a network that consists of 37 neurons, we can predict the network state in the next moment with an average error of less than a quarter of the neurons per instance. Using the machine learning methods, we classify each individual measured neuronal state into the state of solving the task, or the state of reward. Not only that we can classify states with accuracy greater than 95%, but we can also understand the activity of which neurons signal that mouse is solving the task or receiving the reward.

• Keywords: fiziološka mjerenja; neuroni; Isingov model; inverzni problem; strojno učenje; boosting; stablo odluke; ansambli modela; neuronske mreže; SHAP
• Keywords (english): physiological measurements; neurons; Ising’s model; inverse problem; machine learning; boosting; decision tree; model ensembles; neural networks; SHAP
• Language: croatian
• URN:NBN: urn:nbn:hr:217:966036
• Study programme: Title: Physics; specializations in: Research; Course: Research; Study programme type: university; Study level: integrated undergraduate and graduate; Academic / professional title: magistar/magistra fizike (magistar/magistra fizike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2019-01-23 12:25:08