Online Learning Networks#
Build an online-learning-ready recurrent model in three connected steps. First,
ETP operators mark the parameterized computations that participate in online
learning. Next, braintrace.nn layers compose those operators into reusable
network blocks. Finally, hidden states provide the temporal destinations that
the compiler connects to the marked parameter paths.
Choose a foundation#
Mark trainable computation paths with the five public ETP operators and verify their batching, unit, and JAX-transform contracts.
Compose the marked operators through braintrace.nn layers while
preserving operation-based parameter selection and relation boundaries.
Define and initialize the recurrent state that makes a model temporal, then inspect the hidden groups discovered by the compiler.
Recommended sequence#
Step |
Chapter |
Use it to |
|---|---|---|
1 |
Mark parameter operations for online learning and verify their unit and JAX-transform behavior. |
|
2 |
Assemble marked operations into recurrent models without violating parameter-to-hidden relation boundaries. |
|
3 |
Make the model temporal and understand how the compiler forms hidden groups from recurrent state. |
Where to look first#
Unexpected units, batching, or transform behavior: start with Operators for Online Learning.
A weight is excluded or marked non-temporal: inspect the relation-boundary discussion in Neural Network Layers for Online Learning.
State shapes, grouping, initialization, or reset behavior are unclear: use Hidden States for Online Learning.