Master or Semester Project: Comparing RNNs and Point Process Models to uncover Latent Dynamical Mechanisms in Neural Data

Background

Understanding the neural dynamics underlying complex cognitive tasks, such as language processing, is a central goal of neuroscience. At our lab, we are recording population neural activity in the human brain during language tasks. To decipher the underlying dynamics, we currently employ fit point-process models to model the observed spike train data. The fitted models can then be approximated as a continuous dynamical system, facilitating analysis and interpretation. An alternative modeling strategy involves smoothing the raw spike trains and directly fitting a continuous recurrent neural networks (RNNs) to this smoothed data. It remains unclear whether, for a given task and spike train encoding, both approach will uncover the same underlying dynamical mechanisms.

 

Project description

This project aims to comprehensively compare continuous RNNs and point-process models for analyzing neural data recorded during language tasks. We will focus on evaluating the resulting dynamics and assessing how well each approach captures the underlying neural mechanisms.

 

Methodology

You will use existing implementations of both the RNNs and point-process models and fit them to spike train data, either simulated or recorded during cognitive task in humans and non-human primates (e.g. motor control, speech perception). The dynamics of the fitted models will be compared and quantified, focusing on key features like stability, attractors, and dimensionality. Known ground truth dynamics will be evaluated first. If time allows, the same tools will then be used to investigate the reconstructed dynamics in recorded datasets.

 

Requirements

Proficiency in Python is essential.

A solid understanding of dynamical systems theory, including concepts like stability, attractors, and bifurcation analysis.

 

Contact

Prof. Timothée Proix: proix@ini.ethz.ch

 

Starting date and duration

This project is currently available as a semester project or master thesis.

 

Related literature

- Mohammadi et al. (2025). From spiking neuronal networks to interpretable dynamics: a diffusion-approximation framework. BioRxiv https://www.biorxiv.org/content/10.1101/2024.12.17.628339v1

- Sussillo et al. (2013). Opening the Black Box: Low-Dimensional Dynamics in High-Dimensional Recurrent Neural Networks. Neural computation 25(3) https://direct.mit.edu/neco/article/25/3/626-649/7854