Support-CSP 五種新架構與 Shallow 基準
原始 32 通道正常人資料組的正式 10 位受試者；為直接比較既有 Support-CSP，固定選用 Fp1、Fp2、Fz、C3、C4、Pz、O1、O2 八通道。
可信的正向結果：Residual-gated 在三個 all-trial seeds 為 76.32%，Shallow 為 73.91%；paired 差異 +2.41 pp，95% CI [+0.89 pp, +4.04 pp]，8/10 人改善，Wilcoxon p=0.0195。
低 trial 未達標：五組 20+20 抽樣中 Residual 只比 Shallow 高 +0.85 pp，95% CI [-0.70 pp, +2.45 pp]，僅 4/10 人改善。這不是穩定提升，更沒有達到 +5 pp。
第一階段：六模型配對初篩
| 設定 | 模型 | Accuracy | Subject STD | Balanced accuracy | Macro-F1 |
|---|---|---|---|---|---|
| 20+20 | Support-CSP residual gated | 65.00% | 13.07% | 65.00% | 63.77% |
| 20+20 | Support-CSP NewNetV3 late fusion | 63.25% | 11.00% | 63.25% | 61.81% |
| 20+20 | Support-CSP Shallow (baseline) | 63.25% | 12.14% | 63.25% | 61.87% |
| 20+20 | Support-CSP Shallow FiLM | 61.00% | 16.96% | 61.00% | 60.10% |
| 20+20 | Support-CSP multiscale gate | 60.50% | 14.66% | 60.50% | 59.10% |
| 20+20 | Support-CSP token attention | 59.00% | 13.13% | 59.00% | 57.44% |
| All trial | Support-CSP residual gated | 76.61% | 13.92% | 76.59% | 76.48% |
| All trial | Support-CSP Shallow (baseline) | 73.72% | 14.29% | 73.70% | 73.47% |
| All trial | Support-CSP Shallow FiLM | 72.56% | 14.71% | 72.55% | 72.33% |
| All trial | Support-CSP multiscale gate | 71.23% | 13.06% | 71.25% | 70.99% |
| All trial | Support-CSP NewNetV3 late fusion | 71.15% | 13.17% | 71.13% | 70.89% |
| All trial | Support-CSP token attention | 66.35% | 12.74% | 66.35% | 65.95% |
每個模型 100 folds：10 位 × 5 folds × all/固定20+20。六模型共用完全相同的 fold indices、target seed、CSP support/query 與 34 位來源 episodes。
第二階段：Residual 與 Shallow 多次確認
| 設定 | 模型 | Accuracy | Subject STD | Balanced accuracy | Macro-F1 |
|---|---|---|---|---|---|
| 20+20 × 5 subsets | Support-CSP residual gated | 63.40% | 9.04% | 63.40% | 62.00% |
| 20+20 × 5 subsets | Support-CSP Shallow (baseline) | 62.55% | 9.45% | 62.55% | 60.83% |
| All trial × 3 seeds | Support-CSP residual gated | 76.32% | 14.08% | 76.30% | 76.17% |
| All trial × 3 seeds | Support-CSP Shallow (baseline) | 73.91% | 14.42% | 73.91% | 73.73% |
| 設定 | Residual - Shallow | 95% CI low | 95% CI high | 改善人數/10 | Wilcoxon p |
|---|---|---|---|---|---|
| 20+20 × 5 subsets | +0.85 pp | -0.70 pp | +2.45 pp | 4 | 0.367188 |
| All trial × 3 seeds | +2.41 pp | +0.89 pp | +4.04 pp | 8 | 0.019531 |
架構與參數量
| 模型 | 核心方向 | 參數量 |
|---|---|---|
| Support-CSP Shallow (baseline) | 既有 Shallow temporal/spatial convolution、square/log-power pooling 與 CSP late fusion。 | 105,190 |
| Support-CSP NewNetV3 late fusion | NewNetV3 三尺度時間卷積及輕量 channel attention，CSP 在分類前 late fusion。 | 48,294 |
| Support-CSP Shallow FiLM | CSP 產生有界 gamma/beta，條件化 Shallow waveform embedding。 | 125,282 |
| Support-CSP multiscale gate | 25/51/101-sample 三個 log-power 分支，由 CSP 決定尺度權重。 | 28,969 |
| Support-CSP residual gated | 完整空間卷積、三層 dilated depthwise residual blocks、log-power pooling 與 CSP gate。 | 68,410 |
| Support-CSP token attention | Shallow log-power 時間格加 CSP conditioning token，使用兩層 self-attention 融合。 | 194,418 |
Residual 候選沒有放棄 ShallowConvNet 的核心假設：先完整跨通道空間卷積，再以 square/log-power 聚合；新增的是三層 dilation 1/2/4 的 depthwise residual temporal blocks，以及 CSP 控制的有界 feature gate。NewNetV3 與 attention 候選雖能降低來源 loss，正式受試者反而較差。
來源 episodic 預訓練
來源為 sub16-sub49 共 34 位，正式 sub1/sub2/sub3/sub9-sub15 完全未參與。每個 epoch/subject 的 12+12 support 與 32+32 query 對所有架構一致。
單一受試者確認結果
| 受試者 | All Shallow | All Residual | All 差異 | 20+20 Shallow | 20+20 Residual | 20+20 差異 |
|---|---|---|---|---|---|---|
| sub1 | 91.63% | 94.09% | +2.45 pp | 75.00% | 73.50% | -1.50 pp |
| sub10 | 75.00% | 81.17% | +6.17 pp | 65.00% | 65.00% | +0.00 pp |
| sub11 | 92.83% | 94.33% | +1.50 pp | 73.50% | 73.00% | -0.50 pp |
| sub12 | 82.95% | 84.90% | +1.96 pp | 66.50% | 69.00% | +2.50 pp |
| sub13 | 76.00% | 79.83% | +3.83 pp | 61.50% | 66.00% | +4.50 pp |
| sub14 | 56.83% | 59.17% | +2.33 pp | 48.50% | 47.00% | -1.50 pp |
| sub15 | 57.83% | 57.89% | +0.06 pp | 53.00% | 56.50% | +3.50 pp |
| sub2 | 64.50% | 64.02% | -0.49 pp | 58.50% | 55.50% | -3.00 pp |
| sub3 | 85.83% | 84.83% | -1.00 pp | 72.00% | 72.00% | +0.00 pp |
| sub9 | 55.73% | 62.98% | +7.25 pp | 52.00% | 56.50% | +4.50 pp |
資料洩漏與完整性稽核
CSP 只用 outer-train/support labels；validation/test labels 不參與 CSP fit。測試中翻轉所有非 support labels，CSP features 逐值不變。train/val/test 互斥，且每個配對 fold 的三組 index hash 在模型間完全一致。
{
"screen": {
"profile": "screen",
"models": [
"support_csp_shallow",
"support_csp_newnetv3_late",
"support_csp_shallow_film",
"support_csp_multiscale_gate",
"support_csp_residual_gated",
"support_csp_token_attention"
],
"expected_rows_per_model": 100,
"rows_per_model": {
"support_csp_shallow": 100,
"support_csp_newnetv3_late": 100,
"support_csp_shallow_film": 100,
"support_csp_multiscale_gate": 100,
"support_csp_residual_gated": 100,
"support_csp_token_attention": 100
},
"duplicate_keys": 0,
"max_split_hash_variants_across_models": 1,
"finite_metrics": true,
"complete": true
},
"confirmation": {
"profile": "full",
"models": [
"support_csp_shallow",
"support_csp_residual_gated"
],
"expected_rows_per_model": 400,
"rows_per_model": {
"support_csp_shallow": 400,
"support_csp_residual_gated": 400
},
"duplicate_keys": 0,
"max_split_hash_variants_across_models": 1,
"finite_metrics": true,
"complete": true
}
}
檔案
六模型 fold CSV · 確認 fold CSV · 受試者摘要 · paired 統計 · 來源 loss · 程式碼