pp-prop Examples#
The examples/pp_prop/ series is the input/output-factorized online-learning
track. pp_prop is the canonical public name; ES_D_RTRL remains a
backward-compatible alias.
Follow the numbered scripts in order:
01-basics-lif-integrator.py introduces the LIF integrator. API: pp-prop.
02-neurons-alif-dms.py applies an ALIF model to delayed matching-to-sample. API: pp-prop.
03-neurons-gif-working-memory.py demonstrates a GIF working-memory model. API: pp-prop.
04-neurons-coba-ei-rsnn.py uses a conductance-based E/I recurrent SNN. API: SignedWLinear.
05-batching-vmap.py uses per-sample
vmapexecution. API: compile.06-batching-batched.py uses batched primitives. API: compile.
07-vjp-single-step.py demonstrates a single-step VJP. API: pp-prop.
08-vjp-multi-step.py demonstrates a multi-step VJP. API: pp-prop.
09-operator-sparse.py exercises a sparse recurrent operator. API: SparseLinear.
10-operator-lora.py exercises low-rank recurrent weights. API: lora_matmul.
11-operator-conv.py exercises a convolutional SNN. API: Conv2d.
12-classification-neuromorphic.py provides a neuromorphic classification example. API: pp-prop.
13-knob-decay-vs-rank.py compares trace decay with rank. API: pp-prop.
14-knob-vjp-method-contrast.py contrasts VJP methods. API: pp-prop.
Read the pp-prop examples README and the pp-prop tutorial alongside the scripts. The factorization reduces memory but introduces approximation error; its suitability depends on the model dynamics and selected decay or rank.