API Reference

braintrace trains recurrent and spiking neural networks online — forward in time, without backpropagation through time (BPTT). You mark the trainable operations of a model with ETP primitives (for example braintrace.matmul() instead of x @ w); a compiler then discovers how each parameter influences the network’s hidden states and wires up the eligibility traces that carry gradient information forward.

This reference is organized around the four layers of the package, with dependencies pointing strictly downward, plus the ready-made neural-network layers built on top.

Layer	What it does	Reference
Operators	User-facing ETP ops that mark weights for online learning, and the machinery to register your own primitives.	ETP Operators & Core Types, Custom ETP Primitives
Compiler	Walks the JAX jaxpr, identifies ETP primitives by type, and connects each parameter to the hidden states it influences.	Compiler, Executor & Diagnostics
Executor	Runs the forward pass and computes the hidden→weight / hidden→hidden Jacobians the algorithms consume.	Compiler, Executor & Diagnostics
Algorithms	Online-learning orchestrators: the exact D-RTRL / ES-D-RTRL family and the SNN algorithms (EProp, OSTL, OTPE, OTTT, OSTTP).	Online-Learning Algorithms
Layers	Drop-in brainstate.nn-style layers pre-wired through ETP primitives.	Neural-Network Layers

The fastest way in is the one-call braintrace.compile() entry point, documented in Online-Learning Algorithms.

• ETP Operators & Core Types
  • ETP Primitive Operators
  • Input Data
  • Eligibility Trace State
  • Gradient Utilities
  • Errors
• Custom ETP Primitives
  • Registration
  • The four ETP rules
  • Registration example
  • Auto-derived JAX rules
  • The gradient_enabled flag
• Compiler, Executor & Diagnostics
  • Graph Compilation
  • Module Info
  • Hidden Groups
  • Hidden–Parameter–Operation Relations
  • Hidden Perturbation
  • Graph Executor
  • Diagnostics
• Online-Learning Algorithms
  • One-Call Entry Point
  • Base Classes
  • D-RTRL — Parameter Dimension (exact)
  • ES-D-RTRL — Input/Output Dimension (exact)
  • SNN Online-Learning Algorithms
  • Algorithm Comparison
• Neural-Network Layers
  • Linear Layers
  • Convolutional Layers
  • Recurrent Layers
  • Readout Layers