Welcome to TENSION’s documentation!

A Tensorflow (Keras) based Python package for FORCE (First-Order, Reduced and Controlled Error) training chaotic recurrent neural networks.

Note

This project is under active development.

Contents

• Installation
  • Installing from GitHub
  • Installing from PyPI
• Usage
  • Quick Start
  • Layer and Model Class Compatibilities
  • Accessing Key Attributes
    • FORCELayer Classes
    • FORCEModel Classes
  • Creating New Layers
    • Custom State Size
    • Custom Kernel Initialization
    • Custom Build
    • Custom from_weights
    • Creating New Spiking Layers
  • Creating New Model Classes
    • Customizing train_step
    • Customizing force_layer_call
    • Customizing Pseudogradient Updates
  • Callbacks
  • Tests
  • XLA
  • GPU Support
• FORCE Layers
  • Base FORCE Layer
  • Echo State Networks
  • Spiking Networks
• FORCE Model
  • Base FORCE Model
  • Inherited FORCE Model
• Examples
  • FORCE and full-FORCE Training of Echo State Networks on Sum of Sinusoids
  • Training FORCE with Zebrafish Neural Data
  • FORCE Training of Theta Spiking Network on Lorenz Attractor
  • FORCE Training of LIF Spiking Network on GO-NO GO with Random Starts
  • full-FORCE Running Time Comparison Between tension and Original Paper
  • FORCE Running Time Comparison Between tension and JAX Numpy Implementation
  • FORCE Running Time Comparison With and Without XLA
• Notes