8 通道 NewNetV3 與 ShallowConvNet 監督轉移比較
固定 Fp1、Fp2、Fz、C3、C4、Pz、O1、O2；1–40 Hz zero-phase IIR、200 Hz、事件後 0–2 秒、每 trial 每通道 Z-score。無 EA、ASR、ICA、baseline feature branch 或 SSL。
結果判讀
單 seed 篩選的最佳設定為 NewNetV3 scratch（61.54%）。NewNetV3 scratch 相對 Shallow scratch 為 +1.20%，但 paired CI 跨 0；NewNetV3 的 34 人初始化在無 replay 時改變 -3.41%，加入 replay 後改變 -5.33%。source-subject validation 僅約 51–52%，顯示負遷移主要在跨受試者預訓練階段已形成。
組別結果
| setting | n_subjects | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 |
|---|---|---|---|---|---|
| NewNetV3 scratch | 10 | 61.54% | 9.15% | 61.49% | 56.98% |
| Shallow scratch | 10 | 60.34% | 11.28% | 60.36% | 57.64% |
| NewNetV3 transfer, no replay | 10 | 58.13% | 5.08% | 58.11% | 57.52% |
| Shallow transfer, no replay | 10 | 57.90% | 7.70% | 57.89% | 57.05% |
| NewNetV3 supervised transfer | 10 | 56.21% | 6.97% | 56.22% | 55.52% |
| Shallow supervised transfer | 10 | 56.21% | 6.39% | 56.19% | 55.47% |
單一受試者
| subject_id | setting | mean_accuracy | std_accuracy | mean_balanced_accuracy |
|---|---|---|---|---|
| sub1 | NewNetV3 scratch | 65.12% | 10.55% | 64.71% |
| sub1 | NewNetV3 supervised transfer | 49.68% | 9.79% | 49.76% |
| sub1 | NewNetV3 transfer, no replay | 61.01% | 19.35% | 60.90% |
| sub1 | Shallow scratch | 78.37% | 6.90% | 78.29% |
| sub1 | Shallow supervised transfer | 59.13% | 5.68% | 58.98% |
| sub1 | Shallow transfer, no replay | 71.49% | 6.34% | 71.38% |
| sub10 | NewNetV3 scratch | 63.00% | 5.42% | 63.00% |
| sub10 | NewNetV3 supervised transfer | 60.50% | 2.09% | 60.50% |
| sub10 | NewNetV3 transfer, no replay | 58.50% | 2.85% | 58.50% |
| sub10 | Shallow scratch | 60.00% | 5.00% | 60.00% |
| sub10 | Shallow supervised transfer | 68.00% | 6.22% | 68.00% |
| sub10 | Shallow transfer, no replay | 68.00% | 7.37% | 68.00% |
| sub11 | NewNetV3 scratch | 78.50% | 5.18% | 78.50% |
| sub11 | NewNetV3 supervised transfer | 59.50% | 15.25% | 59.50% |
| sub11 | NewNetV3 transfer, no replay | 59.00% | 13.87% | 59.00% |
| sub11 | Shallow scratch | 75.00% | 7.29% | 75.00% |
| sub11 | Shallow supervised transfer | 57.00% | 11.37% | 57.00% |
| sub11 | Shallow transfer, no replay | 59.50% | 14.83% | 59.50% |
| sub12 | NewNetV3 scratch | 61.62% | 11.45% | 61.60% |
| sub12 | NewNetV3 supervised transfer | 65.21% | 9.49% | 65.21% |
| sub12 | NewNetV3 transfer, no replay | 60.71% | 3.05% | 60.74% |
| sub12 | Shallow scratch | 61.76% | 13.90% | 62.00% |
| sub12 | Shallow supervised transfer | 52.72% | 2.80% | 52.67% |
| sub12 | Shallow transfer, no replay | 54.74% | 6.03% | 54.67% |
| sub13 | NewNetV3 scratch | 71.50% | 6.75% | 71.50% |
| sub13 | NewNetV3 supervised transfer | 56.00% | 6.27% | 56.00% |
| sub13 | NewNetV3 transfer, no replay | 57.00% | 11.91% | 57.00% |
| sub13 | Shallow scratch | 64.50% | 7.37% | 64.50% |
| sub13 | Shallow supervised transfer | 61.50% | 6.52% | 61.50% |
| sub13 | Shallow transfer, no replay | 62.50% | 8.84% | 62.50% |
| sub14 | NewNetV3 scratch | 51.00% | 2.85% | 51.00% |
| sub14 | NewNetV3 supervised transfer | 58.00% | 8.91% | 58.00% |
| sub14 | NewNetV3 transfer, no replay | 57.00% | 3.26% | 57.00% |
| sub14 | Shallow scratch | 47.00% | 8.91% | 47.00% |
| sub14 | Shallow supervised transfer | 53.50% | 5.76% | 53.50% |
| sub14 | Shallow transfer, no replay | 52.50% | 1.77% | 52.50% |
| sub15 | NewNetV3 scratch | 53.76% | 5.99% | 53.98% |
| sub15 | NewNetV3 supervised transfer | 52.68% | 8.75% | 52.64% |
| sub15 | NewNetV3 transfer, no replay | 52.68% | 9.10% | 52.67% |
| sub15 | Shallow scratch | 44.85% | 10.12% | 44.81% |
| sub15 | Shallow supervised transfer | 43.79% | 5.84% | 43.79% |
| sub15 | Shallow transfer, no replay | 45.73% | 6.32% | 45.71% |
| sub2 | NewNetV3 scratch | 56.13% | 6.52% | 56.12% |
| sub2 | NewNetV3 supervised transfer | 52.72% | 9.79% | 52.67% |
| sub2 | NewNetV3 transfer, no replay | 53.68% | 10.02% | 53.64% |
| sub2 | Shallow scratch | 49.23% | 10.11% | 49.29% |
| sub2 | Shallow supervised transfer | 52.24% | 6.34% | 52.33% |
| sub2 | Shallow transfer, no replay | 52.73% | 6.32% | 52.86% |
| sub3 | NewNetV3 scratch | 65.00% | 11.99% | 65.00% |
| sub3 | NewNetV3 supervised transfer | 65.00% | 7.07% | 65.00% |
| sub3 | NewNetV3 transfer, no replay | 69.50% | 5.70% | 69.50% |
| sub3 | Shallow scratch | 66.00% | 6.52% | 66.00% |
| sub3 | Shallow supervised transfer | 56.50% | 1.37% | 56.50% |
| sub3 | Shallow transfer, no replay | 56.50% | 2.85% | 56.50% |
| sub9 | NewNetV3 scratch | 49.74% | 2.85% | 49.50% |
| sub9 | NewNetV3 supervised transfer | 42.83% | 10.97% | 42.88% |
| sub9 | NewNetV3 transfer, no replay | 52.21% | 10.37% | 52.19% |
| sub9 | Shallow scratch | 56.72% | 7.16% | 56.71% |
| sub9 | Shallow supervised transfer | 57.71% | 3.10% | 57.62% |
| sub9 | Shallow transfer, no replay | 55.26% | 4.86% | 55.24% |
配對比較
| left | right | mean_difference | bootstrap_ci_low | bootstrap_ci_high | wilcoxon_p | subjects_improved | n_subjects |
|---|---|---|---|---|---|---|---|
| newnetv3/scratch | shallowconvnet/scratch | 1.20% | -3.15% | 4.88% | 0.492188 | 6 | 10 |
| newnetv3/supervised_transfer_no_replay | shallowconvnet/supervised_transfer_no_replay | 0.23% | -4.11% | 4.67% | 0.921875 | 5 | 10 |
| newnetv3/supervised_transfer | shallowconvnet/supervised_transfer | 0.00% | -5.33% | 5.19% | 1.000000 | 6 | 10 |
| newnetv3/supervised_transfer_no_replay | newnetv3/scratch | -3.41% | -8.50% | 0.91% | 0.388672 | 3 | 10 |
| shallowconvnet/supervised_transfer_no_replay | shallowconvnet/scratch | -2.45% | -6.81% | 1.74% | 0.375000 | 4 | 10 |
| newnetv3/supervised_transfer | newnetv3/scratch | -5.33% | -10.56% | -0.34% | 0.164062 | 2 | 10 |
| shallowconvnet/supervised_transfer | shallowconvnet/scratch | -4.13% | -9.96% | 1.25% | 0.275391 | 4 | 10 |
34 人跨受試者預訓練診斷
| model | epoch | validation_accuracy | validation_balanced_accuracy |
|---|---|---|---|
| newnetv3 | 2 | 51.94% | 51.92% |
| shallowconvnet | 12 | 50.75% | 50.72% |
Shallow 在第 12 epoch、NewNetV3 在第 2 epoch 達到最佳 source-subject validation，但兩者 balanced accuracy 都接近 chance。這表示模型能擬合來源受試者，卻沒有形成可直接泛化至未見受試者的左右腳決策邊界。
實驗設計
34 位開發受試者先用真實 left/right label 做 subject-balanced supervised training。正式十位受試者每個 outer fold 只抽 20+20 calibration trials；另比較不使用 replay，以及依 calibration covariance 選五位 source subjects、微調 batch 以 1:1 target/source replay 組成。20+20 內部的 4+4 只用來選 epoch，之後以全部 40 trials從相同初始化重訓。本頁為單 seed 篩選，transfer 未改善，故未追加多 seed。
fold CSV · subject CSV · group CSV · paired CSV · trial manifest · run config