Expand description
Acoustic-event detection head (P64, fail-closed) — AU-LEARNED.
AedHead is a learned acoustic-event classifier that loads its weights through the common
fail-closed loader. With no weights it abstains (BackendUnavailable) — it never emits a
fabricated event. With a (possibly synthetic/test) weight blob present it runs a CPU reference
linear classifier over a caller-supplied feature vector and writes a single low-assurance
event proposal (flagged FLAG_LOW_ASSURANCE | FLAG_REFERENCE_BACKEND), never ground truth.
Blob convention: dims = [num_classes, num_features], row-major data of
num_classes * num_features weights followed by num_classes bias terms. The forward pass is
logit_c = bias_c + Σ_f weight[c,f] * features[f]; the argmax class is proposed. Inference is
the hot path and allocates nothing (streaming argmax over caller buffers).
Structs§
- AedHead
- Learned acoustic-event detection head. Fail-closed:
Absentstate ⇒ inference abstains.