Određivanje dinamike fizikalnog sustava pomoću interpretabilnih algoritama dubokog učenja

master's thesis

• Title (english): Determining the dynamics of a physical system using interpretable deep learning algorithms
• Author: Bartol Pavlović
• Mentor: Davor Horvatić (mentor)
• Committee member: Davor Horvatić (predsjednik povjerenstva)
• Committee member: Matko Glunčić (član povjerenstva)
• Committee member: Petar Žugec (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Physics); Zagreb
• Defense date and country: 2025-01-20, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Physics

Abstract

Ovaj diplomski rad istražuje primjenu interpretabilnih algoritama dubokog učenja za modeliranje dinamike populacije neurona koristeći podatke vremenskih nizova. Neuronske simulacije generirane Brian 2 simulatorom služe kao primarni skup podataka, replicirajući ponašanje međusobno povezane mreže neurona u kontroliranim uvjetima. U istraživanju se primjenjuje novi duboki konvolucijski okvir osmišljen za multivarijatne vremenske nizove, s naglaskom na interpretabilnost bez gubitka točnosti... More predikcije. To se postiže arhitekturom sposobnom za hvatanje aktivnosti pojedinačnih neurona i kolektivne dinamike, kao i pružanje uvida u temeljne mehanizme koji upravljaju tim ponašanjima. Prepoznavanjem najbitnijih značajki filtriranjem signala od šuma i aproksimacijom interakcija zadanim funkcijama, model spaja fleksibilnost dubokog učenja i interpretabilnost klasičnih regresija. Fokusirajući se na Izhikievichev model neurona, rad istražuje sposobnost modeliranja sustava vremenskih nizova i interpretabilnost modela. Rezultati pokazuju da ovakav interpretabilni pristup dubokog učenja može učinkovito analizirati i predvidjeti neuronsku dinamiku, povezujući područja strojnog učenja, neuroznanosti i fizike. Također, razvijene su nove ideje za prilagođavanje izlaza modela promatranom sustavu. Less

Abstract (english)

This thesis explores the application of interpretable deep learning algorithms for modeling the dynamics of neuron populations using time-series data. Neural simulations generated with the Brian 2 simulator serve as the primary dataset, replicating the behavior of interconnected neural networks under controlled conditions. The study employs a novel deep convolutional framework for multivariate time series, emphasizing interpretability without sacrificing predictive accuracy. This is... More achieved through an architecture capable of capturing the activity of individual neurons and collective dynamics while providing insights into the fundamental mechanisms governing these behaviors. By identifying key features, filtering signals from noise, and approximating interactions with predefined functions, the model combines the flexibility of deep learning with the interpretability of classical regression techniques. Focusing on the Izhikevich neuron model, the work examines the capacity of time-series systems to model and interpret the dynamics. The results demonstrate that such an interpretable deep learning approach can effectively analyze and predict neural dynamics, bridging the fields of machine learning, neuroscience, and physics. Additionally, new ideas were developed for adapting the model's outputs to the observed system. Less

• Keywords: dinamika neurona; Izhikievichev model; predviđanje multivarijatnih vremenskih nizova; interpretabilno duboko učenje; interpretabilnost modela; mehanistička interpretabilnost
• Keywords (english): neuron dynamics; Izhikevich model; multivariate time series forecasting; interpretable deep learning; model interpretability; mechanistic interpretability
• Language: croatian
• URN:NBN: urn:nbn:hr:217:479192
• 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: 2025-01-17 10:01:48