Neural Manifold Visualization

This repo provides Python examples for visualizing the manifolds.

Overview

python neural_manifolds.py

The visualization consists of two main components:

1. A spike raster plot showing the firing patterns of three individual neurons
2. A 3D trajectory plot showing the neural manifold (the path traced by the combined neural activity)

python umap_example.py Example of how to use UMAP for visualization of latent space.

1. A neural network learns to classify MNIST digits
2. Visualize the latent space using UMAP during training

Installation

1. Create and activate a new conda environment:

conda create env create --name nengo --file env.yml
conda activate nengo

2. Or install the required Python packages by yourself:

pip install numpy==1.26.3
pip install nengo==4.0.0
pip install matplotlib

3. Install ffmpeg (for MP4 export):

On Ubuntu/Debian:

sudo apt-get install ffmpeg

On Conda:

conda install ffmpeg

Technical Details

Neural Network Configuration

• Uses 3 Leaky Integrate-and-Fire (LIF) neurons
• Each neuron has a different encoding vector aligned with one axis
• Firing rates: 200-400 Hz
• Membrane time constant (tau_rc): 0.02s
• Synaptic filtering: 0.01s on input, 0.05s on output

Input Signal

• Smooth circular trajectory in 3D space
• Frequency: 0.5 Hz
• Amplitude: 0.8
• Third dimension phase-shifted by π/4

Visualization Features

• Color-coded spike trains for each neuron
• Gradient-colored trajectory showing time progression
• Rotating 3D view for better spatial perception
• 1-second sliding window for spike raster
• Continuous looping animation

Outputs

The script generates two animation files:

1. neural_manifold.mp4: High-quality video format
   • Higher quality
   • Smaller file size
   • Requires video player support
2. neural_manifold.gif: Animated GIF format
   • More widely compatible
   • Larger file size
   • Can be viewed in any web browser

Customization

Key parameters that can be modified:

• max_rates: Adjust neuron firing rates
• tau_rc: Change membrane time constant
• window: Modify the spike raster window size
• freq: Adjust trajectory frequency
• amplitude: Change trajectory size
• FPS and bitrate for animation export