🧠 Reconstructing Hidden Neural Dynamics with Autoencoders

A Python-based project exploring how effectively autoencoders—a foundational deep learning architecture—can reconstruct latent neural states and system dynamics under varying degrees of partial observability. Given that neural recordings are often noisy and incomplete, this project demonstrates how autoencoders can robustly infer underlying neural patterns, offering valuable insights for neural signal analysis and clinical interpretation. Beyond research, this work has real-world relevance: improving the design of brain–computer interfaces (BCIs) by enhancing the temporal decoding of neural activity, which is crucial for precise control in neuroprosthetics and brain-driven digital systems.

• Training:
  • Simulated cortical activity using the Jansen-Rit Neural Mass Model with six state variables over 10,000 time steps.
  • Built a fully connected autoencoder (128–64–3–64–128) in PyTorch, trained with Adam optimiser over 50 epochs.
  • Used a custom MSE loss to handle partial observability, where input dimensions varied (1–6 states).
• Testing:
  • Assessed performance under three input observability conditions: All six states (full observability) vs Three membrane potentials only (partial) vs Only pyramidal neuron potential (extreme partial)
  • Measured Pearson correlation between predicted and true states.
  • Analysed latent space dynamics across conditions.
• Findings:
  • Autoencoders effectively reconstructed full dynamics under complete observability.
  • Under partial conditions, inhibitory interneuron activity was more robustly inferred than excitatory neurons, suggesting asymmetric recoverability depending on network dynamics.
  • Latent space degraded from chaotic to linear with limited input, indicating significant information loss
  • Results suggest promise for using autoencoders in uncovering hidden neural dynamics, with room for improvement using deeper or variational architectures.

The repository consists of the Jupyter-Notebook script (CAS2425_FinalProject_RongDing.ipynb) and the resulting project report (Final_Project_RD.pdf).