eprintid: 37427
rev_number: 15
eprint_status: archive
userid: 9314
dir: disk0/00/03/74/27
datestamp: 2025-10-14 11:54:25
lastmod: 2025-10-14 11:54:48
status_changed: 2025-10-14 11:54:25
type: doctoralThesis
metadata_visibility: show
creators_name: Thurm, Max Ingo
title: Reconstructing Neural Dynamics Underlying Cognitive Flexibility Using Parameter-Evolving RNNs
subjects: ddc-000
subjects: ddc-530
subjects: ddc-570
divisions: i-140001
adv_faculty: af-14
cterms_swd: Dynamical Systems Reconstruction
cterms_swd: RNN
cterms_swd: rule Learning
cterms_swd: decision making
abstract: Understanding the dynamic principles that enable the brain to flexibly adapt behavior in changing environments remains a central challenge in neuroscience. In
this thesis, I address this question through the lens of dynamical systems reconstruction. I use a reconstruction method, specifically targeted for non-autonomous
neural dynamics from multiple single-unit recordings in the rodent medial prefrontal
cortex (mPFC) during a probabilistic rule-learning task. To this end, I employ a
parameter-evolving piecewise-linear recurrent neural network (pePLRNN), which
explicitly incorporates time-dependent changes in the underlying dynamical system
(DS). This approach enables the reconstruction of non-autonomous DSs from nonstationary data to characterize of how the neural dynamics evolve across learning.
The approach was first validated on benchmark systems and task-trained RNNs,
where it successfully reconstructed the underlying DS. When trained on the hidden
state trajectories of RNNs solving artificial rule-learning tasks, the pePLRNN uncovered the dynamic mechanisms by which these networks implemented the learning
process.
Applied to electrophysiological recordings from the mPFC of rats, the model successfully reconstructed the non-stationary neural dynamics underlying rule learning.
The trained model-generated neural trajectories that exhibited the same decoding
properties as the original data. Change points (CP) detected in model-generated trajectories aligned with those observed in the recorded activity. Simulations of neural
trajectories under experimental conditions reproduced the behavioral distributions
of animals for both rule types.
Analyzing the trained pePLRNN as a functional surrogate model revealed that
both rules were implemented via a single stimulus-dependent attracting region that
guided neural transients toward the correct decision. During learning, this attracting region, along with the trial-specific parameters and latent neural trajectories,
exhibited abrupt changes that preceded the behavioral change point.
This work establishes a principled framework for reconstructing non-autonomous
DS directly from empirical data and demonstrates how their analysis as surrogate
models can reveal dynamic principles underlying the neural computations supporting
cognitive flexibility.
date: 2025
id_scheme: DOI
id_number: 10.11588/heidok.00037427
own_urn: urn:nbn:de:bsz:16-heidok-374271
date_accepted: 2025-07-18
advisor: HASH(0x561f65dcab50)
language: eng
bibsort: THURMMAXINRECONSTRUC20251008
full_text_status: public
place_of_pub: Heidelberg
citation:   Thurm, Max Ingo  (2025) Reconstructing Neural Dynamics Underlying Cognitive Flexibility Using Parameter-Evolving RNNs.  [Dissertation]     
document_url: https://archiv.ub.uni-heidelberg.de/volltextserver/37427/1/Max_Ingo_Thurm_Thesis_ubpup.pdf