pyjae: Joint Autoencoders (JAE) for Neural Signal Denoising

pyjae (pronounced "pie-jay") is the official implementation of joint autoencoder models for denoising high-dimensional neural population recordings and recovering the low-dimensional manifold underneath, based on Altan et al. (2021).

Installation

pip install pyjae

Or with uv:

uv add pyjae

For development from source:

git clone https://github.com/egealtan/pyjae.git
cd pyjae
uv sync            # installs the package with the dev extras

Quick start

from pyjae import JAE, simulate_neural_data

# Simulate a 6D nonlinear manifold observed on 64 noisy channels
clean, noisy, info = simulate_neural_data(
    n_samples=400, n_channels=64, n_timepoints=96,
    latent_dim=6, snr_db=10.0, nonlinear=True, alpha=3.0, seed=0,
)

# Train the channel-split model and denoise
model = JAE(latent_dim=6)          # backend="jae1" by default
model.fit(noisy, epochs=200)
denoised = model.denoise(noisy)
print("VAF:", model.score(clean, denoised))

Use the JEPA backend for representation learning plus a denoising readout:

model = JAE(latent_dim=32, backend="jepa", patch_len=8, d_model=64)
model.fit(noisy, epochs=150)
denoised = model.denoise(noisy)

Package layout

Requirements

• Python >= 3.10
• PyTorch >= 2.0, NumPy, scikit-learn, SciPy

Citation

@article{altan2021jae,
  title={Estimating the dimensionality of the manifold underlying multi-electrode neural recordings},
  author={Altan, Ege and Solla, Sara A. and Miller, Lee E. and Perreault, Eric J.},
  journal={PLOS Computational Biology},
  year={2021},
  volume={17},
  number={11},
  pages={e1008591},
  doi={10.1371/journal.pcbi.1008591}
}

License

MIT License. See LICENSE.