Tutorial 81: SNN Transfer Learning

Pretrain, save, load, freeze, and fine-tune an SNN checkpoint. The workflow is designed for reproducible transfer experiments where the checkpoint carries both weights and enough architecture metadata to reject incompatible reloads.

Save and Load Checkpoints

Python

import numpy as np
from sc_neurocore.transfer import SNNCheckpoint, load_checkpoint, save_checkpoint

rng = np.random.default_rng(42)

model_weights = [
    rng.standard_normal((256, 784)).astype(np.float64) * 0.05,
    rng.standard_normal((10, 256)).astype(np.float64) * 0.1,
]

checkpoint = SNNCheckpoint(
    weights=model_weights,
    layer_names=["hidden", "output"],
    layer_sizes=[(784, 256), (256, 10)],
    neuron_types=["LIF", "LIF"],
    metadata={
        "task": "mnist",
        "accuracy": 0.972,
        "timesteps": 25,
        "beta": 0.9,
    },
)
save_checkpoint(checkpoint, "mnist_snn")

loaded = load_checkpoint("mnist_snn")
print(loaded.n_layers, loaded.total_params, loaded.metadata["task"])

The loader checks the JSON sidecar, rejects unexpected archive entries, disables pickle while opening the .npz, and verifies that each weight matrix shape is (output_features, input_features) for the stored layer-size pair.

Freeze and Fine-Tune

Freeze feature-extraction layers and keep the output head trainable:

Python

from sc_neurocore.transfer import TransferConfig, apply_transfer_config

config = TransferConfig(
    freeze_until=0,
    lr_backbone=0.0,
    lr_head=0.001,
)

loaded, per_layer_lr = apply_transfer_config(loaded, config)
print(loaded.frozen_layers)
print(per_layer_lr)

freeze_until accepts either a zero-based layer index or a layer name. Unknown layer names and out-of-range indices raise ValueError instead of silently leaving the schedule unchanged.

Gradual Unfreezing

Python

from sc_neurocore.transfer import TransferConfig, apply_transfer_config, unfreeze_layers

phase_1 = TransferConfig(freeze_until=0, lr_backbone=0.0, lr_head=0.001)
loaded, phase_1_rates = apply_transfer_config(loaded, phase_1)

unfreeze_layers(loaded, layer_names=["hidden"])
phase_2 = TransferConfig(freeze_until=-1, lr_backbone=0.0001, lr_head=0.001)
loaded, phase_2_rates = apply_transfer_config(loaded, phase_2)

The helpers mutate the checkpoint and return it for call chaining. Save the checkpoint again after each phase when you need an auditable recovery point.

Transfer Workflow

1. Pretrain on the source task.
2. Save a validated SNNCheckpoint.
3. Load the checkpoint for the target task.
4. Replace or reinitialize task-specific head weights outside the checkpoint helper.
5. Freeze the transferred backbone and assign per-layer learning rates.
6. Fine-tune, unfreeze gradually when needed, then save the adapted checkpoint.

What Transfers in SNNs

Component	What Transfers	Why It Helps
Synaptic weights	Feature detectors and temporal filters	Preserves learned spike-response structure.
Layer metadata	Layer order, sizes, and neuron labels	Prevents incompatible checkpoint reuse.
Frozen-layer markers	Fine-tuning schedule state	Makes transfer phases reproducible.
Experiment metadata	Task, accuracy, timestep, or provenance fields	Keeps downstream analysis tied to the saved state.

Evidence

benchmarks/results/bench_transfer.json records local, non-isolated regression evidence for the Python API and polyglot mirrors. The 2026-06-27 run reports:

Check	Result
Python checkpoint roundtrip	100 calls in 0.154779 s, 646.081 calls/s
Python apply_transfer_config	100 calls in 0.009153 s, 10925.32 calls/s
Rust checkpoint and fine-tune mirrors	compile and unit tests pass
Julia transfer validation	pass
Mojo checkpoint and fine-tune kernels	pass

The benchmark artifact is regression evidence only and sets production_benchmark_claim to false.

References

• Yosinski et al. (2014). "How transferable are features in deep neural networks?" NeurIPS 2014.
• Dampfhoffer et al. (2022). "Are SNNs really more energy-efficient than ANNs? An in-depth hardware-aware study." IEEE TETCI.