Loading...
Files
Date
relationships.isAuthorOf
Publisher
Philipps-Universität Marburg
item.page.supervisor-of-thesis
Abstract
The human brain is frequently characterized as a "black box" in the context of comprehending sensory processing and neural optimization. The relationship between input stimuli and the brain's neuronal responses, such as spike rate, tends to remain obscured until it's measured through experiments. This challenge is further compounded by the intrinsic variability of neural responses, which are subject to biological noise, resulting in varying responses even when the same stimulus is repeatedly presented (Goris et al., 2014; Stein et al., 2005).
This thesis focuses on the dorsal visual pathway of primates, a region of the brain that plays a crucial role in interpreting self-motion not only from visual but also from vestibular, tactile and even auditory cues (NHP physiology: Bremmer et al., 2002b; Chen et al., 2011; Schlack et al., 2002; Human fMRI: Bremmer et al., 2001; Huk et al., 2002; Krala et al., 2019; Rosenblum et al., 2023). Non-human primate (NHP) experiments have shown that neurons within this pathway exhibit sensitivity to specific directions of self-motion and typically demonstrate peak activation when the stimuli align with their preferred heading. The directional tuning properties have been the subject of extensive research, focusing especially on the medial superior temporal (MST) area (Bremmer et al., 2010; Duffy and Wurtz, 1995; Lappe et al., 1996; Maunsell and Van Essen, 1983) and the ventral intraparietal area (VIP) (Bremmer et al., 2002b, 2002a; Kaminiarz et al., 2014).
Understanding which stimuli evoke the strongest average responses from such neurons can offer critical insights into the functional organization of the underlying brain area (Dayan and Abbott, 2005). In line with this, tuning curve analysis and reverse correlation techniques have proven instrumental in characterizing preferred stimulus features (Kaminiarz et al., 2014; Kriegeskorte and Wei, 2021; Ringach and Shapley, 2004).
To explore stimulus-response dynamics for visual self-motion stimuli more systematically, I developed a closed-loop optimization framework that adaptively generates visual stimuli to maximize neural activation in real time. At the heart of this approach is a Variational Autoencoder (VAE), which captures the essential structure of complex self-motion stimuli in a compact latent space (Kingma and Welling, 2022). Using the VAE’s decoder, the system can generate new stimuli from any point within this latent space, enabling a focused and efficient search across possible visual inputs.
For optimization, I used a Covariance Matrix Adaptation Evolution Strategy (CMA-ES), an evolutionary algorithm well-suited for high-dimensional, noisy problems (Hansen and Ostermeier, 2001). Here, the optimization objective was defined by neural output, such as spike rate or model-based neural activation, and aimed to find the latent parameters that would drive these outputs as high as possible.
To validate the framework, I applied it to a 3D-ResNet model (He et al., 2016), using its internal filters as stand-ins for actual neural responses. The system successfully identified input patterns that activated specific filters near their tuning peaks, offering the possibility for valuable insights into which features most strongly drive neural-like behavior (Olah et al., 2018), proving the approach suitable for identifying optimal stimuli for in-vivo studies.
Review
Metadata
Contributors
Supervisor:
Dates
Created: 2025Issued: 2025-11-06Updated: 2025-11-06
Faculty
Fachbereich Physik
Language
eng
Data types
DoctoralThesis
Keywords
Neuronale Verarbeitung im dorsalen visuellen PfadNeuronale TuningmechanismenNeural processing in the dorsal visual pathwayAI-based optimization processesNeural tuning mechanismsKI-basierte Optimierungsprozesse
DFG-subjects
ReizverarbeitungNeurowissenschaftenMaschinelles Lernen
DDC-Numbers
500
show more
Kreß, Alexander (0000-0003-0646-5952): Optimization of neural response in the primate dorsal visual pathway. : Philipps-Universität Marburg 2025-11-06. DOI: https://doi.org/10.17192/z2025.0657.