Saruhan Mete Gürbüz

  

Home

Hi, I'm Saruhan Mete Gürbüz, an M1 student in the Computational Neuroscience & AI track of the Cog-SUP master's program jointly run by Sorbonne Université and Université Paris Cité. I'm interested in understanding how neural systems and learning algorithms implement computation, and how these mechanisms give rise to behavior in agents. To this end, I aim to use and develop state of the art Machine Learning methods, and mainly interested in probabilistic frameworks such as Variational Autoencoders. I am currently doing an internship at INM/INSERM(Computational Brain Team) under Joao Barbosa and Philippe Domenech, where I am working on a foundation model for decoding monoamine concentrations from FSCV time series data. Below are selected projects I have worked on or am currently developing.

Contact: [email protected]

Selected Projects

Motor Cortex Population Dynamics Analysis

Neuroscience jPCA Tensor Decomposition Dec 2025 – Present · Personal project

Built analysis pipeline for MC_RTT dataset to study rotational dynamics in motor cortex populations. Implemented noise injection experiments to test robustness of jPCA-extracted manifolds. Applied sliceTCA to demix covariability classes before trajectory analysis.

More details
  • Analysis pipeline for studying rotational dynamics in motor cortex populations.
  • Noise injection experiments to test manifold robustness.
  • Applied tensor decomposition (sliceTCA) to demix covariability classes.

Bayesian Agent Learning & Inference

Bayesian Inference Cognitive Modeling Sept 2025 – Dec 2025 · Course Project

Inferred Bayesian agent priors using analytical and maximum likelihood methods. Developed algorithms to infer learning rules governing belief updates.

More details
  • Course project in Bayesian modeling of brain and behavior.
  • Analytical and maximum likelihood inference methods.
  • Algorithm development for learning rule inference.

GINE–VGAE for FDG-PET Alzheimer's Diagnosis

Graph ML Neuroimaging PyTorch Oct 2024 – Jun 2025 · 93.8% acc, 0.937 F1

Edge-informed variational graph autoencoder for PET-based Alzheimer’s classification, published at IEEE IPTA 2025.

More details
  • Used ADNI FDG-PET Data, implemented graphs where each ROI is a node.
  • Implemented graph preprocessing(pruned unimportant edges via NBS)
  • Key design choice: edge features included in encoding layer of VGAE.
  • Working on interpretability of the results and extending the framework.

Recommender System from Web-scraped Reviews

NLP Recommenders Data pipeline Oct 2023 – Jan 2025 · TÜBİTAK project

Built a sentiment-analysis-driven recommendation model using web-scraped review data.

More details
  • Scraping + cleaning + labeling strategy.
  • Modeling choices and evaluation setup.