Surrogate Gradients

Surrogate gradients provide differentiable approximations for non-differentiable spiking neuron operations. This section collects practical guides that show how to work with the surrogate module, from first principles to advanced customization.

• 01_basics.ipynb walks through the rationale behind surrogate gradients, demonstrates the built-in surrogate functions, and highlights common training patterns for spiking networks.

• 02_customizing.ipynb dives deeper into extending the library, illustrating how to register custom surrogate kernels, combine them with existing layers, and validate their gradients.

• 03_meta_learning.ipynb explores meta-learning techniques for optimizing surrogate hyperparameters, showing how to take derivatives with respect to surrogate parameters and adapt them during training for improved performance.

• Tutorial 1: Surrogate Gradients for Spiking Neural Networks - Basics
• Tutorial 2: Customizing Surrogate Gradients for SNNs
• Tutorial 3: Meta-Learning with Surrogate Hyperparameters