選擇紀錄
| stage | experiment | subjects | accuracy |
|---|---|---|---|
| attribution | attr_legacy_20_eegnet_s20260713 | 10 | 59.25% ± 13.23% |
| attribution | attr_legacy_20_eegnet_s20260714 | 10 | 59.00% ± 9.14% |
| attribution | attr_legacy_20_eegnet_s20260715 | 10 | 57.75% ± 10.50% |
| attribution | attr_legacy_20_eegnet_s20260716 | 10 | 57.75% ± 7.31% |
| attribution | attr_legacy_20_eegnet_s20260717 | 10 | 53.75% ± 8.76% |
| attribution | attr_legacy_all_eegnet | 10 | 65.86% ± 10.83% |
| attribution | attr_raw_20_eegnet_s20260713 | 10 | 57.75% ± 10.77% |
| attribution | attr_raw_20_eegnet_s20260714 | 10 | 57.50% ± 12.36% |
| attribution | attr_raw_20_eegnet_s20260715 | 10 | 55.50% ± 6.10% |
| attribution | attr_raw_20_eegnet_s20260716 | 10 | 52.50% ± 7.55% |
| attribution | attr_raw_20_eegnet_s20260717 | 10 | 53.75% ± 5.80% |
| attribution | attr_raw_all_eegnet | 10 | 66.95% ± 12.48% |
| attribution | attr_raw_all_mibiot | 10 | 58.50% ± 10.50% |
| attribution | attr_raw_all_shallow | 10 | 72.57% ± 16.35% |
| ea_ssl | final_scratch | 10 | 69.58% ± 15.32% |
| ea_ssl | final_scratch_ea | 10 | 71.67% ± 15.53% |
| ea_ssl | final_ssl | 10 | 57.28% ± 8.49% |
| ea_ssl | final_ssl_ea | 10 | 62.32% ± 10.76% |
| preprocessing_filter | filter_zero_1_40 | 10 | 72.53% ± 17.45% |
| preprocessing_filter | filter_zero_8_30 | 10 | 58.18% ± 8.03% |
| preprocessing_normalization | norm_global | 10 | 71.47% ± 16.60% |
| preprocessing_normalization | norm_p95 | 10 | 72.82% ± 18.00% |
| preprocessing_normalization | norm_per_channel | 10 | 71.09% ± 18.21% |
| preprocessing_phase | phase_causal_1_40 | 10 | 72.78% ± 15.98% |
| preprocessing_phase | phase_zero_1_40 | 10 | 71.82% ± 16.16% |
| temporal_spectral | ts_fusion | 10 | 69.41% ± 16.35% |
| temporal_spectral | ts_spectral | 10 | 52.36% ± 4.15% |
| temporal_spectral | ts_temporal | 10 | 68.10% ± 15.92% |
{
"stage1": {
"winner": "attr_raw_all_shallow",
"model": "shallow"
},
"stage2_normalization": {
"winner": "norm_p95",
"normalization": "p95"
},
"stage2_filter": {
"winner": "filter_zero_1_40",
"filter": "1_40"
},
"stage2_phase": {
"winner": "phase_causal_1_40",
"phase": "causal",
"data_mode": "raw_causal_1_40"
},
"stage3": {
"winner": "ts_fusion",
"model": "fusion"
},
"stage4": {
"initial_seed_best_two": [
"final_scratch_ea",
"final_scratch"
]
}
}
一、公平歸因
先比較相同 folds、增強與 normalization 的 EEGNet、ShallowConvNet、MI-BIOT，再用五組固定 20+20 抽樣 seed 量化歷史結果受抽樣運氣影響的程度。
二、前處理消融
依序選 normalization、MI 常用的 8–30 Hz 候選與 1–40 Hz，再比較 zero-phase 與可部署的 causal SOS。每一步只帶勝出設定往下，避免用同一批資料搜尋大型全因子組合。
三、TemporalSpectralNet
Temporal branch 保留多尺度瞬時波形；spectral branch 使用 Mu、low beta、high beta 的 task log-power、baseline dB change 與 C3/C4 側化；fusion 檢查兩者是否互補。
四、EA 與 VICReg SSL
Euclidean Alignment 依 subject/session 平均 covariance 對齊；正式受試者每 fold 僅由 train/calibration trials 估計矩陣。30 位未使用受試者做 subject-balanced VICReg，並以 gradient reversal 降低受試者身分資訊。Scratch、Scratch+EA、SSL、SSL+EA 使用相同 encoder、folds 與微調設定。SSL 預訓練因 GB10/PyTorch 2.13 alpha 的 BF16/FP16 kernel 不穩定改用 FP32；下游為 BF16 autocast，FFT 固定 FP32。
| stage | experiment | subjects | accuracy |
|---|---|---|---|
| ea_ssl | final_scratch | 10 | 69.58% ± 15.32% |
| ea_ssl | final_scratch_ea | 10 | 71.67% ± 15.53% |
| ea_ssl | final_ssl | 10 | 57.28% ± 8.49% |
| ea_ssl | final_ssl_ea | 10 | 62.32% ± 10.76% |
Paired statistics
| comparison | mode_a | mode_b | n_subjects | mean_delta_pp | bootstrap_ci_low_pp | bootstrap_ci_high_pp | wilcoxon_p_raw | a_better_subjects | wilcoxon_p_holm |
|---|---|---|---|---|---|---|---|---|---|
| Scratch + EA vs Scratch | final_scratch_ea | final_scratch | 10 | 2.09 | -0.84 | 5.12 | 0.2324 | 8 | 0.2324 |
| SSL vs Scratch | final_ssl | final_scratch | 10 | -12.30 | -17.30 | -6.87 | 0.0059 | 1 | 0.0293 |
| SSL + EA vs Scratch + EA | final_ssl_ea | final_scratch_ea | 10 | -9.35 | -15.74 | -3.29 | 0.0243 | 1 | 0.0971 |
| SSL + EA vs Scratch | final_ssl_ea | final_scratch | 10 | -7.25 | -13.05 | -2.10 | 0.0273 | 2 | 0.0971 |
| Scratch + EA vs Scratch (3-seed aggregate) | final_scratch_ea_multiseed | final_scratch_multiseed | 10 | 3.50 | 0.73 | 6.40 | 0.0488 | 8 | 0.0977 |
結果判斷
三 seed 平均 Scratch 為 68.91%，Scratch + EA 為 72.41%；配對差異 +3.50 pp，95% CI [0.73, 6.40]，改善受試者 8/10，達到預先設定的 +3 pp、CI 下界大於 0、7/10 同方向標準。Wilcoxon raw p=0.0488，Holm 校正 p=0.0977。
解讀界線：三 seed EA 效果達到預設 effect-size/CI/方向一致性門檻，但五個 planned comparisons 經 Holm 校正後未達 0.05，應視為有支持的探索性結果，而非已完成獨立 confirmatory validation。62.75% NewNetV3 為歷史手動結果，原始程式不存在，因此不對它做顯著性推論。ASR+ICA 只用來回答歸因問題，不列入 online 候選。所有負結果均保留。