# results_fbcnet_spk

**FBCNet-style parallel band Conv DirectNet 10s + SPK/vlaai_ws warm-start**.
Frozen FIR banks [(1.0, 3.0), (3.0, 5.0), (5.0, 7.0), (7.0, 9.0)] Hz → per-band Spatial(id)+Temporal(delta) → shared VLAAI × soft-mix. SPK remap `bb.*`→`bb.bb.*` (strict=False). **NOT** residual FIR/Morlet→vlaai; **NOT** random-init TSception/WaveNet.

Swap-test delta=0.0077. Prep `eaP_n5_t60`, sw2, Fuglsang, max 4 ep×250, batch 64. HMM p_stay 0.994 scale 0.8; also hp30.

Code: `fbcnet_spk_lodo.py` / `decoders.FBCNetSPK`. Checkpoints: `models/fbcnet_spk/held-*.pt`, `/root/disk_3tb/eeg-thao/models/all5_direct_clf_w10b3_ptSPK_fbcnet_aug-sw2.pt`. JSON: `/root/disk_3tb/eeg-thao/results/direct_clf__eaP_n5_t60_w10b3_ptSPK_fbcnet_aug-sw2.json`.

**ABANDONED**: loss 0.7034 still ~chance after 50 steps.

Root cause: SPK remaps into shared VLAAI, but each branch sees **bandpassed** EEG. Soft-mix of four band views ≠ raw-EEG path SPK was trained on (residual FIR/Morlet starts as `raw+0` identity; this does not). SwapΔ=0.0077 passed; learning did not.

## AASD Acc vs tip 53.53

No external score (chance abort). Δ n/a vs 53.53.

## Acc>60?

**NO** (aborted before LODO/ext).

## NEXT IDEA

Close multi-band→vlaai family (parallel soft-mix **and** residual FIR/Morlet). Prefer Acc-tip evidence transforms that do not retrain the EEG front-end (domain emission / conf residual), or a non-bandpass frame prior that preserves an identity path to SPK weights at step 0.
