34人 Hybrid SSL→跨受試者 Bridge：正常人 20+20
結果判斷：Hybrid SSL→bridge 未達 67%。擴大解凍範圍後仍未超過 Hybrid scratch，且所有 Hybrid 模式低於 Shallow；同 34 人先 SSL 再 supervised bridge 在方法上可行，但本次產生負遷移，主要瓶頸是 encoder/objective 與強烈跨受試者 domain shift。
Group accuracy
| model | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 |
|---|---|---|---|---|
| Shallow scratch | 64.98% | 13.25% | 64.99% | 64.65% |
| SSL only + PEFT | 51.54% | 5.00% | 51.55% | 50.94% |
| Supervised bridge + PEFT | 53.58% | 4.27% | 53.58% | 52.97% |
| SSL → bridge + PEFT | 52.43% | 3.27% | 52.44% | 51.41% |
單一受試者
| subject_id | model | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 |
|---|---|---|---|---|---|
| sub1 | Supervised bridge + PEFT | 53.98% | 6.46% | 53.87% | 52.94% |
| sub10 | Supervised bridge + PEFT | 49.33% | 6.01% | 49.33% | 48.85% |
| sub11 | Supervised bridge + PEFT | 58.50% | 5.16% | 58.50% | 57.97% |
| sub12 | Supervised bridge + PEFT | 55.54% | 9.14% | 55.57% | 54.63% |
| sub13 | Supervised bridge + PEFT | 61.83% | 6.51% | 61.83% | 61.25% |
| sub14 | Supervised bridge + PEFT | 54.50% | 6.42% | 54.50% | 54.14% |
| sub15 | Supervised bridge + PEFT | 52.42% | 4.46% | 52.50% | 51.81% |
| sub2 | Supervised bridge + PEFT | 48.43% | 5.42% | 48.40% | 48.13% |
| sub3 | Supervised bridge + PEFT | 49.33% | 7.04% | 49.33% | 48.74% |
| sub9 | Supervised bridge + PEFT | 51.93% | 7.18% | 51.90% | 51.25% |
| sub1 | Shallow scratch | 83.40% | 4.47% | 83.43% | 83.32% |
| sub10 | Shallow scratch | 61.17% | 8.23% | 61.17% | 61.03% |
| sub11 | Shallow scratch | 82.17% | 7.43% | 82.17% | 82.12% |
| sub12 | Shallow scratch | 67.98% | 7.30% | 67.98% | 67.86% |
| sub13 | Shallow scratch | 73.50% | 6.04% | 73.50% | 73.37% |
| sub14 | Shallow scratch | 48.33% | 8.27% | 48.33% | 47.59% |
| sub15 | Shallow scratch | 52.06% | 8.85% | 52.07% | 51.39% |
| sub2 | Shallow scratch | 53.36% | 6.83% | 53.32% | 52.88% |
| sub3 | Shallow scratch | 75.50% | 6.14% | 75.50% | 75.33% |
| sub9 | Shallow scratch | 52.37% | 7.18% | 52.40% | 51.63% |
| sub1 | SSL → bridge + PEFT | 55.36% | 7.47% | 55.37% | 54.45% |
| sub10 | SSL → bridge + PEFT | 48.33% | 8.80% | 48.33% | 47.46% |
| sub11 | SSL → bridge + PEFT | 52.33% | 5.78% | 52.33% | 51.37% |
| sub12 | SSL → bridge + PEFT | 52.89% | 8.65% | 52.94% | 52.48% |
| sub13 | SSL → bridge + PEFT | 60.17% | 9.47% | 60.17% | 57.91% |
| sub14 | SSL → bridge + PEFT | 51.83% | 5.63% | 51.83% | 50.18% |
| sub15 | SSL → bridge + PEFT | 50.94% | 8.85% | 50.98% | 49.49% |
| sub2 | SSL → bridge + PEFT | 51.19% | 7.52% | 51.23% | 50.89% |
| sub3 | SSL → bridge + PEFT | 50.50% | 5.84% | 50.50% | 49.95% |
| sub9 | SSL → bridge + PEFT | 50.75% | 5.74% | 50.74% | 49.91% |
| sub1 | SSL only + PEFT | 53.38% | 7.93% | 53.40% | 52.90% |
| sub10 | SSL only + PEFT | 50.00% | 7.07% | 50.00% | 49.52% |
| sub11 | SSL only + PEFT | 52.67% | 8.10% | 52.67% | 51.81% |
| sub12 | SSL only + PEFT | 47.27% | 6.42% | 47.27% | 46.89% |
| sub13 | SSL only + PEFT | 64.33% | 8.32% | 64.33% | 63.97% |
| sub14 | SSL only + PEFT | 50.17% | 3.95% | 50.17% | 49.25% |
| sub15 | SSL only + PEFT | 52.59% | 9.46% | 52.63% | 51.62% |
| sub2 | SSL only + PEFT | 47.28% | 3.78% | 47.33% | 46.70% |
| sub3 | SSL only + PEFT | 49.50% | 6.14% | 49.50% | 48.93% |
| sub9 | SSL only + PEFT | 48.24% | 7.99% | 48.21% | 47.83% |
跨受試者 Bridge validation
| mode | mean_validation_bacc | std_validation_bacc | median_best_epoch |
|---|---|---|---|
| bridge_only | 52.94% | 1.03% | 12.0 |
| ssl_bridge | 52.79% | 1.19% | 10.0 |
Paired statistics
| comparison | mean_delta_pp | bootstrap_ci_low_pp | bootstrap_ci_high_pp | wilcoxon_p | left_better_subjects |
|---|---|---|---|---|---|
| SSL → bridge + PEFT vs Shallow scratch | -12.5533 | -19.7634 | -5.4987 | 0.0137 | 1 |
| SSL → bridge + PEFT vs Supervised bridge + PEFT | -1.1508 | -2.6785 | 0.2870 | 0.2324 | 3 |
| SSL → bridge + PEFT vs SSL only + PEFT | 0.8874 | -0.8431 | 2.5809 | 0.3750 | 6 |
| Supervised bridge + PEFT vs Shallow scratch | -11.4024 | -18.5073 | -4.5097 | 0.0195 | 2 |
解凍範圍診斷
Adapter-only 只更新 9,903 參數；partial 更新 118,175 參數並解凍 spatial convolution 與最後一個 Transformer block；full 更新全部 243,047 參數。
| model | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 |
|---|---|---|---|---|
| Bridge partial | 57.11% | 5.90% | 57.12% | 56.72% |
| Hybrid scratch full | 58.05% | 6.28% | 58.05% | 55.69% |
| SSL → bridge partial | 51.87% | 3.48% | 51.86% | 50.05% |
| SSL full | 53.73% | 4.83% | 53.75% | 52.91% |
| comparison | mean_delta_pp | bootstrap_ci_low_pp | bootstrap_ci_high_pp | wilcoxon_p | left_better_subjects |
|---|---|---|---|---|---|
| Bridge partial vs Hybrid scratch full | -0.9333 | -2.8187 | 0.7980 | 0.5566 | 4 |
| SSL full vs Hybrid scratch full | -4.3187 | -8.1134 | -0.2553 | 0.1055 | 3 |
| SSL → bridge partial vs Bridge partial | -5.2455 | -8.3557 | -2.2955 | 0.0098 | 2 |
| Bridge partial vs Shallow scratch | -7.8695 | -13.5180 | -2.5565 | 0.0371 | 3 |
Latent SSL
fold results · adaptation folds · trial manifest · group summary · subject summary · paired statistics · adaptation statistics · bridge CV · 34-subject manifest · run config · code