08 — Neural nets, CNNs, LSTMs, baselines

flowchart LR
    IN["beat in"] --> CNN["CNN: pattern detectors\nslide along, fire on match"]
    CNN --> BIG["stacked layers\nsmall patterns build big ones"]
    BIG --> OUT["N/V/S/F/Q vote"]

flowchart LR
    S1["step 1"] --> S2["step 2 + memory notebook"] --> S3["step 3 + notebook"]

Model	Reads	Strength	Weakness
1D CNN	raw signal	fast, position-proof QRS detectors	no rhythm memory
2D CNN	scalogram image	time-frequency patterns	needs the scalogram built
LSTM	step by step + memory	rhythm across beats	slow, fussy to train

Baselines (../research/plan.md Phase 2) are the honesty mechanism: 1D-CNN, then CNN+LSTM. Mamba earns its place only by beating both on macro F1 and V recall.

Next: 09-ssm-mamba.md