8通道 ConvMAE SSL 與 ShallowConvNet 左右腳分類

固定 Fp1、Fp2、Fz、C3、C4、Pz、O1、O2；zero-phase SOS IIR；200 Hz；task 0–2 秒；baseline -0.5–0 秒。

結果判斷： ShallowConvNet 達到 70%，表示固定 8 通道與兩秒窗口具有足夠分類資訊；ConvMAE scratch 與 SSL 都較低，因此問題主要在架構 inductive bias 與 SSL objective，不能再解釋成資料本身的 accuracy ceiling。

鎖定設定

Filter: 1_40; normalization: per_channel; capacity: medium; SSL full: recon_transform /full

Group accuracy

Group accuracy
model n_subjects mean_accuracy std_accuracy mean_balanced_accuracy mean_macro_f1
ShallowConvNet 10 72.67% 18.14% 72.65% 71.47%
ConvMAE scratch 10 68.32% 14.74% 68.33% 67.03%
SSL full fine-tune 10 67.51% 14.48% 67.49% 66.59%

單一受試者

Subject accuracy
subject_id model mean_accuracy std_accuracy mean_balanced_accuracy mean_macro_f1
sub1 ConvMAE scratch 85.23% 5.30% 85.20% 85.16%
sub10 ConvMAE scratch 69.00% 7.12% 69.00% 68.25%
sub11 ConvMAE scratch 87.50% 6.55% 87.50% 87.39%
sub12 ConvMAE scratch 74.95% 5.50% 74.94% 74.70%
sub13 ConvMAE scratch 73.50% 4.10% 73.50% 73.05%
sub14 ConvMAE scratch 47.83% 5.89% 47.83% 43.90%
sub15 ConvMAE scratch 51.26% 6.17% 51.25% 47.73%
sub2 ConvMAE scratch 56.41% 9.69% 56.51% 53.97%
sub3 ConvMAE scratch 82.50% 10.22% 82.50% 82.33%
sub9 ConvMAE scratch 55.04% 8.58% 55.06% 53.81%
sub1 ShallowConvNet 94.56% 3.24% 94.57% 94.55%
sub10 ShallowConvNet 79.50% 4.11% 79.50% 79.44%
sub11 ShallowConvNet 95.00% 0.00% 95.00% 95.00%
sub12 ShallowConvNet 84.57% 1.17% 84.60% 84.49%
sub13 ShallowConvNet 74.00% 5.76% 74.00% 73.78%
sub14 ShallowConvNet 49.00% 6.52% 49.00% 44.73%
sub15 ShallowConvNet 51.74% 4.47% 51.67% 49.19%
sub2 ShallowConvNet 60.56% 3.84% 60.60% 59.87%
sub3 ShallowConvNet 86.50% 5.48% 86.50% 86.46%
sub9 ShallowConvNet 51.27% 9.26% 51.12% 47.24%
sub1 SSL full fine-tune 85.10% 6.08% 85.12% 85.01%
sub10 SSL full fine-tune 70.50% 8.51% 70.50% 70.17%
sub11 SSL full fine-tune 88.17% 4.27% 88.17% 88.14%
sub12 SSL full fine-tune 73.58% 9.00% 73.60% 73.45%
sub13 SSL full fine-tune 72.17% 6.61% 72.17% 71.88%
sub14 SSL full fine-tune 51.67% 7.18% 51.67% 49.52%
sub15 SSL full fine-tune 51.94% 10.00% 51.93% 49.41%
sub2 SSL full fine-tune 56.43% 7.83% 56.30% 55.85%
sub3 SSL full fine-tune 77.33% 4.86% 77.33% 77.24%
sub9 SSL full fine-tune 48.24% 3.99% 48.11% 45.20%

20+20 少量 Trial 測試

Few-shot 判斷： SSL full 使用每類 20 個 calibration trials 時為 59.89%，ShallowConvNet 為 65.26%。SSL 未維持 67%，因此目前不能宣稱少量 trial 已達到原本 all-trial 水準。

每個 outer fold 僅從 calibration pool 抽取左腳 20、右腳 20，40 筆全數用於訓練；完整 outer test fold 不參與抽樣或 epoch 選擇。epoch 預算由既有 all-trial inner-validation 結果換算成相同 median optimizer-update budget，兩模型共用相同子集並重複三個固定 seed。

20+20 group accuracy
model all_trial_mean all_trial_std fewshot_mean fewshot_std fewshot_minus_all_pp
ShallowConvNet 72.67% 18.14% 65.26% 12.73% -7.41
SSL full fine-tune 67.51% 14.48% 59.89% 8.38% -7.62

20+20 單一受試者

20+20 subject accuracy
subject_id model mean_accuracy std_accuracy mean_balanced_accuracy mean_macro_f1 n_fold_seed_results
sub1 ShallowConvNet 83.43% 7.81% 83.45% 83.38% 15
sub10 ShallowConvNet 67.00% 5.76% 67.00% 66.78% 15
sub11 ShallowConvNet 78.00% 6.96% 78.00% 77.87% 15
sub12 ShallowConvNet 67.63% 8.15% 67.63% 67.43% 15
sub13 ShallowConvNet 74.50% 5.19% 74.50% 74.30% 15
sub14 ShallowConvNet 48.00% 9.12% 48.00% 47.40% 15
sub15 ShallowConvNet 50.58% 7.02% 50.58% 49.82% 15
sub2 ShallowConvNet 53.84% 6.67% 53.82% 53.64% 15
sub3 ShallowConvNet 75.50% 6.07% 75.50% 75.35% 15
sub9 ShallowConvNet 54.11% 7.33% 54.17% 53.77% 15
sub1 SSL full fine-tune 65.39% 8.65% 65.35% 64.53% 15
sub10 SSL full fine-tune 61.33% 5.74% 61.33% 60.94% 15
sub11 SSL full fine-tune 73.67% 6.87% 73.67% 73.35% 15
sub12 SSL full fine-tune 61.18% 9.60% 61.15% 60.31% 15
sub13 SSL full fine-tune 65.67% 7.23% 65.67% 65.38% 15
sub14 SSL full fine-tune 50.50% 7.21% 50.50% 49.12% 15
sub15 SSL full fine-tune 51.42% 9.85% 51.40% 50.43% 15
sub2 SSL full fine-tune 52.47% 9.67% 52.47% 52.09% 15
sub3 SSL full fine-tune 67.50% 6.88% 67.50% 67.09% 15
sub9 SSL full fine-tune 49.77% 7.28% 49.76% 48.81% 15

20+20 Paired statistics

comparison n_subjects mean_delta_pp bootstrap_ci_low_pp bootstrap_ci_high_pp wilcoxon_p ssl_better_subjects
SSL full vs ShallowConvNet (20+20 all-train fixed budget) 10 -5.368699 -8.945996 -2.21501 0.013672 2
model comparison n_subjects mean_delta_pp bootstrap_ci_low_pp bootstrap_ci_high_pp wilcoxon_p fewshot_better_subjects
ShallowConvNet 20+20 minus all-trial 10 -7.4118 -11.6894 -3.1138 0.0195 2
SSL full fine-tune 20+20 minus all-trial 10 -7.6220 -11.6757 -3.7642 0.0098 1

few-shot fold results · trial manifest · protocol config · subject summary · group summary · paired model statistics · few-shot vs all-trial statistics

Validation holdout 敏感度分析

若嚴格把 20+20 再拆成每類 16 train、4 validation，實際訓練只剩 32 筆，而且 validation 指標每錯一筆就跳動 12.5 percentage points。此結果用來評估 early-stopping 不穩定性，不作主要 few-shot 結論。

model mean_accuracy std_accuracy mean_balanced_accuracy mean_macro_f1
ShallowConvNet 61.37% 10.42% 61.38% 57.79%
SSL full fine-tune 55.81% 5.54% 55.81% 53.20%

holdout fold results · holdout subject summary · holdout group summary

Preprocessing screen

filter normalization val_balanced_accuracy
1_40 global 78.52%
1_40 per_channel 79.27%
8_30 global 66.83%
8_30 per_channel 63.80%

模型容量

size val_balanced_accuracy
compact 77.40%
medium 78.30%

SSL 與微調策略

objective strategy val_balanced_accuracy
recon full 77.32%
recon linear 65.15%
recon staged 75.08%
recon_transform full 77.52%
recon_transform linear 64.54%
recon_transform staged 75.52%

Paired statistics

comparison mode_a mode_b n_subjects mean_delta_pp bootstrap_ci_low_pp bootstrap_ci_high_pp wilcoxon_p a_better_subjects
SSL full vs matched scratch ssl_full convmae_scratch 10 -0.810163 -2.715854 0.913841 0.695312 5
SSL full vs ShallowConvNet ssl_full shallowconvnet 10 -5.158537 -7.815884 -2.365447 0.013672 2
ConvMAE scratch vs ShallowConvNet convmae_scratch shallowconvnet 10 -4.348374 -7.103262 -1.497154 0.013672 1

預訓練

Reconstruction pretrainingTransformation pretraining

fold results · subject summary · group summary · paired statistics · screening · fine-tuning ablation · selection · run config · reconstruction history · transformation history · reproducible code