Tutorial 61: SNN Explainability — Why Did It Fire?

Three methods to explain SNN decisions at the spike level.

Spike Attribution

Trace which input spikes contributed most to the output:

from sc_neurocore.explain import SpikeAttributor

attr = SpikeAttributor(decay=0.9)
result = attr.attribute(input_spikes, weights, output_neuron=0)

# Top 5 most important spikes
for t, n, score in result.top_k(5):
    print(f"  t={t}, neuron={n}: importance={score:.3f}")

Perturbation Saliency

Remove each spike and measure output change:

from sc_neurocore.explain import TemporalSaliency

sal = TemporalSaliency(run_fn=my_model)
result = sal.explain(input_spikes, output_neuron=0)
print(result.summary())

Causal Importance

Silence each neuron and measure classification impact:

from sc_neurocore.explain import CausalImportance

ci = CausalImportance(run_fn=my_model)
result = ci.explain(input_spikes, output_neuron=0)
# result.importance_map[0] = per-neuron importance scores

Comparison

Method	What It Shows	Speed	Granularity
Attribution	Weight-path importance	Fast (one pass)	Per spike per timestep
Saliency	Output sensitivity	Slow (N perturbations)	Per spike per timestep
Causal	Neuron-level impact	Medium (N interventions)	Per neuron