Looma „GPS“ süsteemi uurimine: Masinõppimise kaudu hipokampuse funktsiooni tundma õppimine
Files
Date
2015
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
2014. aasta Nobeli preemia füsioloogias said Dr. John M. O’Keefe, Dr. May-Britt Moser ja Dr. Edvard I teatud kindlate rakkude avastamise eest ajus, mis vastutavad ruumi- ja suunataju eest. Need avastused võimaldavad arvata, et aju loob sisemise kaardi ümbritsevast keskkonnast. See aitab meil ära tunda tuttavaid kohti ja ruumis hästi orienteeruda. Antud magistritöös kasutasime me roti “GPS” süsteemi tundma õppimiseks arvutuslikku lähenemist. Konkreetsemalt võrdlesime, kui hästi suudavad erinevad masinõppe algoritmid ette ennustada roti asukohta, saades sisendiks ainult tema hipokampuses toimuva neuronaalse aktiivsuse. Võrreldud meetodite seas olid juhumets (random forest), tugivektorklassifitseerijad (support vector machine, SVM), lähima naabri meetod (nearest neighbor) ja mõningad hajusa lineaarse regressiooni algoritmid. Neuronitest mõõdetud elektrofüsioloogilised andmed olid pärit Buxsaki laborist New Yorgis. Keskendusime roti hipokampuse neuraalsele aktiivsusele - aju osale, kus on senistes uurimustöödes enamik koharakke tuvastatud.
Esimese sammuna jagasime ala, kus rott eksperimendi ajal viibis, neljaks väiksemaks tsooniks. Seejärel üritasime ennustada, missuguses alas katsealune loom mingil suvalisel ajahetkel asus. Leidsime, et juhumets andis parima ennustustäpsuse, milleks oli 57.8% ja mis on oluliselt suurem juhusliku valiku tõenäosusest. Sellegipoolest oli mõnedes katseala regioonides tugivektorklassfitseerija mõnikord parem kui juhumets. Järgmise sammuna tegime asukoha identifitseerimise veelgi raskemaks ja jagasime eksperimentaalala 16 väiksemaks tsooniks. Juhumets ja SVM saavutasid tugevalt statisiliselt olulised tulemused, vastavalt 38% ja 37% (juhusliku ennustuse täpsus oleks olnud umbes 11%). Mõlema probleemülesande puhul kasutasime me ka lähima naabri algoritmi, aga selle täpsus oli võrreldes eelmainitud meetoditega märgatavalt väiksem. Kuna roti asukoht on pidev muutuja, siis me proovisime käsitleda seda ka pideva ennustuse probleemina. Suurem osa regressiooni algoritme, mida selles töös analüüsitakse (kantregressioon (ridge regression), lassoregressioon (lasso regression), elastne võrk (elastic net)), andsid juhuslikule ennustustäpsusele lähedasi tulemusi. Ainult juhumets andis pideva ennustuse probleemi puhul teistest meetoditest oluliselt parema täpsuse.
Seejärel analüüsisime me andmeid, mis olid salvestatud eksperimendist, kus rotid olid treenitud valima vasakut või paremat suunda number 8 kujulises labürindis, olles samal ajal ise jooksurattal. Nende mõõtmistulemuste puhul teostasime me esimese sammuna andmetele mõõdete vähendamise (dimensionality reduction), et visualiseerida muutusi andmetes otsuse langetamise hetkel. Muuhulgas identifitseerisime ja tõime joonistel välja ka episoodirakud - neuronid, mis on rohkem aktiivsed kindlal ajal antud ülesande jooksul. Episoodirakud võivad kaasa aidata aja tajumisel ja episoodilise mälu loomisel. Samuti visualiseerisime neuronaalseid trajektoore otsuse langetamise ajal, et ette aimata, millise otsuse loom vastu võtab.
Kokkuvõtteks andis roti asukoha ennustamisel algoritmidest täpseimaid tulemusi juhumets. See võib muuhulgas näidata seda, et informatsioon roti asukoha kohta sisaldub mitte-lineaarses neuraalses aktiivsuses, mida lineaarregressiooni meetodid ei olnud võimelised tuvastama. Edasises uurimistöös plaanime me dekodeerida roti asukohta, kasutades meetodeid, mis on sarnasemad aju enda mehhanismidele. Neurovõrgud (neural networks) on laialt levinud masinõppe meetod, mis sarnaselt juhumetsadega suudab ära tunda mitte-lineaarseid mustreid. Selles töös loodud andmetöötluskonveiereid (data processing pipeline), mis tegelevad üsnagi keerulise andmete eeltöötluse, tunnuste eraldamise ja andmestiku visualiseerimisega, panevad tulevikuks tugeva aluse hipokampuse dünaamika uurimisele TÜ arvutusliku neuroteaduse töögrupis.
The 2014 Nobel prize in Physiology was awarded to Dr. John M. O’Keefe, Dr. May-Britt Moser and Dr. Edvard I for discovering particular cells in the brain that provide the sense of place and navigation. These discoveries suggest that the brain creates internal map-like representation of the environment which helps us recognize familiar places and navigate well. In this thesis, we used a computational approach to study the animal "GPS" system. In particular, we set to compare how well different machine learning algorithms are able to predict a rat's position just based on its hippocampal neural activity. Methods compared included Random Forest, Support Vector Machines, k-Nearest Neighbors, and several sparse linear regression algorithms. Data was obtained multi-neuron electrophysiological data recorded from the Buzsaki lab in New York, and we focus on the activity of rat hippocampus, the brain region where most the place cells have been identified. In a first step, we divided the experimental arena into 4 blocks and tried to classify in which one of those blocks the rat was at a given time. In this case, we found that Random Forest gave the best accuracy which was 57.8%, well beyond the chance level. However, in some particular regions of the arena, Support Vector Machine was sometimes better than Random Forest. For the next step, we made the classification problem even harder by dividing the arena into 16 blocks. Random Forest and SVM produced highly significant results with 38% and 37% accuracy respectively (random classifier accuracy would be approximately ~11%). We also used K-Nearest Neighbors for both classification problems but its accuracy was less in both cases than the above mentioned algorithms. Since the rat position is a continuous variable we also considered the continuous prediction problem. Most regression algorithms we analyzed (Ridge Regression, LASSO, Elastic Net) provided results near chance level while Random Forest outperformed the algorithms and gave the best results in this case. Furthermore, we analysed data recorded from an experiment where rats were trained to choose left or right direction in a 8-shaped maze while they were running in a wheel. In this case we perform a dimensionality reduction of the neuronal data to visualize its dynamics during the decision time. We also identified and provided plots of episodic cells (neurons who are more active at particular times in the task) which might contribute to the sense of time and create episodic memory. Also, we visualized neuronal trajectories while animal makes decisions in order to predict its future decision. In conclusion, from the algorithms we analysed Random Forest gave the best accuracy while predicting a rat's location. This might also indicate that the information about rat location is contained in non-linear patterns of neuronal activity, which linear regression methods were unable to extract. In future research we plan to decode a rat position using a method more similar to the brain own mechanisms such as neural networks, which as Random Forest can detect non-linear patterns. More generally, the pipelines developed during this thesis to handle the complex pre-processing, feature extraction, and visualization of the dataset will set the basis for future studies on hippocampal dynamics by the group of computational neuroscience in the University of Tartu.
The 2014 Nobel prize in Physiology was awarded to Dr. John M. O’Keefe, Dr. May-Britt Moser and Dr. Edvard I for discovering particular cells in the brain that provide the sense of place and navigation. These discoveries suggest that the brain creates internal map-like representation of the environment which helps us recognize familiar places and navigate well. In this thesis, we used a computational approach to study the animal "GPS" system. In particular, we set to compare how well different machine learning algorithms are able to predict a rat's position just based on its hippocampal neural activity. Methods compared included Random Forest, Support Vector Machines, k-Nearest Neighbors, and several sparse linear regression algorithms. Data was obtained multi-neuron electrophysiological data recorded from the Buzsaki lab in New York, and we focus on the activity of rat hippocampus, the brain region where most the place cells have been identified. In a first step, we divided the experimental arena into 4 blocks and tried to classify in which one of those blocks the rat was at a given time. In this case, we found that Random Forest gave the best accuracy which was 57.8%, well beyond the chance level. However, in some particular regions of the arena, Support Vector Machine was sometimes better than Random Forest. For the next step, we made the classification problem even harder by dividing the arena into 16 blocks. Random Forest and SVM produced highly significant results with 38% and 37% accuracy respectively (random classifier accuracy would be approximately ~11%). We also used K-Nearest Neighbors for both classification problems but its accuracy was less in both cases than the above mentioned algorithms. Since the rat position is a continuous variable we also considered the continuous prediction problem. Most regression algorithms we analyzed (Ridge Regression, LASSO, Elastic Net) provided results near chance level while Random Forest outperformed the algorithms and gave the best results in this case. Furthermore, we analysed data recorded from an experiment where rats were trained to choose left or right direction in a 8-shaped maze while they were running in a wheel. In this case we perform a dimensionality reduction of the neuronal data to visualize its dynamics during the decision time. We also identified and provided plots of episodic cells (neurons who are more active at particular times in the task) which might contribute to the sense of time and create episodic memory. Also, we visualized neuronal trajectories while animal makes decisions in order to predict its future decision. In conclusion, from the algorithms we analysed Random Forest gave the best accuracy while predicting a rat's location. This might also indicate that the information about rat location is contained in non-linear patterns of neuronal activity, which linear regression methods were unable to extract. In future research we plan to decode a rat position using a method more similar to the brain own mechanisms such as neural networks, which as Random Forest can detect non-linear patterns. More generally, the pipelines developed during this thesis to handle the complex pre-processing, feature extraction, and visualization of the dataset will set the basis for future studies on hippocampal dynamics by the group of computational neuroscience in the University of Tartu.