Label Budget and Adaptation

RQ3. How do label budget, adaptation strategy, and same-architecture Scratch controls change the evidence for pretraining-attributable benefit?

Full Fine-Tuning

Pretrained FM parameters and downstream heads are optimized under each label budget.

Lower scoreHigher scoreUnsupported
Dataset · task · splitCBraModCSBrainCodeBrainEEGMambaEEGPTLEADNeuroRVQREVE
1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%
HGDLeft/Right/Hand/Foot/TongueCS0.3020.3310.220.3530.5530.490.580.190.2630.2730.330.450.5030.6130.3010.050.3210.3720.6020.5720.6420.2130.2820.260.220.480.440.53-0.070.000.050.080.140.180.320.06-0.060.120.200.400.490.590.00--0.00---0.040.150.3020.4910.6710.6310.711
Meng2019Cursor Control + RestCS0.2120.080.090.120.100.080.040.1930.140.130.080.200.230.22-0.030.120.2030.200.160.160.190.2310.2330.170.2430.240.240.2730.070.010.090.130.160.210.26-0.010.2420.170.2920.3330.3430.3920.120.180.242-0.3920.352-0.070.3110.3610.3610.4110.4010.521
OpenBMILeft/Right Hand MICS0.5030.5330.520.550.530.540.520.5610.5720.610.570.600.610.650.380.480.500.500.510.520.500.480.490.6320.5930.630.660.670.5020.440.540.450.510.610.6730.480.530.6410.6220.7330.7820.8220.490.651--0.7520.763-0.450.450.6230.6810.8310.8210.861
PhysioNet-MIHand vs Feet MICS0.16-0.070.2430.1830.240.3820.3820.4920.0320.2620.140.3120.3430.3730.27-0.000.080.140.170.250.310.5010.000.3810.2420.2830.260.290.000.0230.000.000.090.070.140.07-0.040.000.000.150.200.260.00-0.000.00---0.3530.0510.130.2610.3710.4210.421
EEGMatMental Arithmetic vs RestCS0.500.6310.50T0.500.650.7610.8410.67T0.500.460.56T0.580.720.7730.500.500.50T0.500.7320.7430.750.500.500.50T0.500.570.660.650.500.500.490.6810.600.700.730.400.420.470.500.7710.7420.7920.500.500.50T0.500.500.500.500.67T0.380.470.56T0.6830.680.71
TUSLBackground Slowing DetectionCS0.500.500.9310.8110.6220.650.7730.500.500.6430.7730.570.620.621.00T0.500.600.7820.580.7720.7910.500.500.500.610.570.7130.650.500.500.500.500.550.500.580.500.500.510.570.6230.570.7820.500.500.500.500.500.500.501.00T0.500.7020.750.8210.7710.77
ERP-BCIP300 Target DetectionCS0.000.000.2120.3610.4410.5310.5610.000.000.1130.090.3420.4820.5420.1310.1010.2310.1830.250.380.4830.000.000.000.2020.2630.4130.460.0620.000.000.000.000.000.00-0.07-0.040.000.100.150.220.320.000.000.000.000.000.000.00-------
Kaggle-INRIAP300 Feedback CorrectnessCS0.500.500.500.410.510.530.5930.500.440.510.430.500.5820.510.5820.440.5530.500.500.500.500.330.6520.550.480.490.560.590.500.500.530.500.500.490.500.7510.7110.6020.480.6210.560.560.380.500.460.5220.6220.5630.6520.500.440.6010.5210.6230.6410.681
SEEDEmotion RecognitionCS0.1730.2620.2920.3810.190.240.090.2220.2430.230.3030.200.3420.2720.160.090.2930.180.230.240.230.130.170.230.260.2430.2830.2730.010.170.130.150.210.210.210.110.100.220.250.230.240.230.110.11--0.262-0.220.3910.3810.3710.3520.4110.4110.381

Frozen Encoder

The pretrained backbone is fixed, so the matrix isolates head-only adaptation under scarce labels.

Dataset · task · splitCBraModCSBrainCodeBrainEEGMambaEEGPTLEADNeuroRVQREVE
1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%
HGDLeft/Right/Hand/Foot/TongueCS0.090.090.140.1930.4110.3030.4030.1910.2520.1530.120.240.170.320.050.1530.1810.2020.3330.3710.4220.090.2610.140.2410.3920.3620.4610.000.000.000.000.000.040.020.163-0.01-0.020.060.040.110.190.000.000.000.000.000.000.000.1720.050.1620.130.180.160.18
Meng2019Cursor Control + RestCS-0.000.160.090.180.190.090.02-0.000.140.100.150.060.170.130.0330.2420.1030.1930.2230.2330.2330.2010.1830.2020.2710.2510.2320.2920.000.000.000.000.000.000.000.112-0.040.020.040.110.180.200.000.030.020.010.040.050.040.000.2410.2410.2320.2520.3110.331
OpenBMILeft/Right Hand MICS0.480.5220.480.510.500.500.520.5030.490.520.550.540.580.580.400.480.500.540.510.550.580.400.520.5720.6120.5920.6420.6610.420.480.490.430.510.500.520.480.5310.5430.6210.6210.6610.6620.5510.500.5810.500.520.520.520.5120.5230.460.5630.5630.5930.593
PhysioNet-MIHand vs Feet MICS0.070.000.1430.070.2130.200.210.3520.0610.110.1720.2610.2430.2810.1930.000.1720.1030.190.3010.2330.5010.000.3410.2210.2320.2520.2620.000.000.010.00-0.010.000.000.17-0.06-0.080.05-0.00-0.000.110.000.000.000.000.000.000.00-0.40-0.200.000.000.090.150.08
EEGMatMental Arithmetic vs RestCS0.500.500.500.500.500.500.500.8310.500.500.500.540.6710.7020.500.7310.7010.5710.6320.6520.7410.500.500.500.500.5930.6130.6730.500.500.500.500.500.500.500.400.420.410.480.490.500.610.500.500.500.500.500.500.500.500.460.500.500.6310.580.61
TUSLBackground Slowing DetectionCS0.500.500.690.64T0.520.520.530.500.500.710.64T0.500.690.681.00T0.330.9010.6910.6030.7620.7420.500.500.500.610.590.7230.660.500.500.500.500.500.500.500.500.500.73T0.570.7220.520.7330.500.500.500.500.500.500.501.00T0.8310.73T0.570.8410.7710.781
ERP-BCIP300 Target DetectionCS0.00T0.00T0.1720.1730.2520.2830.2330.00T0.00T0.2010.3110.1930.3320.382-0.08-0.010.0930.080.000.250.150.00T0.00T0.000.2020.3210.4010.441-0.01-0.000.000.000.000.000.00-0.05-0.030.000.000.000.000.000.00T0.00T0.000.000.000.000.00-------
Kaggle-INRIAP300 Feedback CorrectnessCS0.50T0.500.5430.410.5130.520.510.420.360.5520.5430.490.5230.5620.380.57T0.500.400.5510.510.510.420.6510.450.490.500.5320.5430.50T0.500.500.490.500.500.500.50T0.57T0.520.5520.490.510.510.50T0.500.500.500.500.500.500.380.530.6110.6310.5520.5310.581
SEEDEmotion RecognitionCS0.2610.2030.240.3210.230.210.040.1730.2510.2920.240.2630.200.220.130.2020.230.210.240.2830.2430.160.180.2930.2730.2720.3120.2720.030.070.070.140.130.170.15-0.010.060.150.210.190.200.200.000.04-0.000.010.100.090.130.2220.070.3010.3020.3210.3610.311

Same-Architecture Scratch

The same FM architectures are trained from random initialization, isolating architecture from pretraining.

Dataset · task · splitCBraModCSBrainCodeBrainEEGMambaLEADNeuroRVQREVE
1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%
HGDLeft/Right/Hand/Foot/TongueCS0.080.201-0.3020.390.320.482---0.053-0.330.21----0.2530.4610.431-0.131--0.200.4420.392-0.08-0.0330.1320.140.4330.370.433-------0.1020.1220.3010.3410.400.3830.521
Meng2019Cursor Control + RestCS0.1720.040.180.150.140.150.000.2010.1820.2130.170.130.190.200.100.183-0.2130.230.190.210.1430.2510.18-0.243--0.000.160.2810.2320.2720.2820.302-0.04-0.000.130.060.180.2030.2330.120.160.2120.2910.3310.3710.391
OpenBMILeft/Right Hand MICS0.5020.6030.5830.520.540.520.510.440.480.560.5630.56-0.570.270.6020.540.5620.530.5430.50-0.6310.582-0.583-0.6530.4830.420.500.530.6210.7320.771-------0.5310.520.6110.6410.6220.7410.742
PhysioNet-MIHand vs Feet MICS0.5210.1410.3910.2310.1920.2020.250.2120.062-0.030.030.1130.160.220.000.000.0630.000.000.170.2630.083-0.020.202-0.1930.2720.000.00-0.020.000.000.000.17-------0.000.000.2020.0930.3310.3310.331
EEGMatMental Arithmetic vs RestCS0.5030.440.500.5120.6410.6420.711-0.5610.5910.6110.5230.560.650.5030.5030.500.500.500.7510.702-0.41-0.500.572-0.6630.5030.420.500.500.500.6030.560.5030.5030.500.500.500.500.500.5030.5030.500.500.500.500.56
TUSLBackground Slowing DetectionCS0.500.500.6620.5820.5330.6220.550.500.500.6430.6210.5520.6210.5920.500.500.7110.500.500.6230.6210.500.50--0.500.51-0.000.500.410.5630.5810.480.550.500.500.500.500.500.500.500.500.500.500.500.500.540.573
ERP-BCIP300 Target DetectionCS0.1310.0730.1620.050.2230.3230.452-0.180.2510.100.0830.3210.4010.4810.000.000.2410.2310.2920.3620.4230.000.1820.1030.102---0.000.000.000.000.120.260.290.000.000.000.000.000.000.00-------
Kaggle-INRIAP300 Feedback CorrectnessCS0.7510.500.5330.5410.5530.520.550.250.6220.591--0.530.550.5030.500.500.500.500.500.50-0.50---0.5830.540.5030.440.530.480.6010.6610.6120.460.6410.470.5420.550.530.5930.5030.5530.5520.5230.5820.6320.661
SEEDEmotion RecognitionCS0.2020.2610.1730.2420.190.210.110.2410.1930.2310.210.21-0.1830.1230.110.1920.261-0.263-0.100.262--0.2720.26-0.07-0.060.070.200.2330.2720.272--------0.09-0.000.080.2230.3210.2810.331

Supervised Baselines

Task-supervised baselines are shown at the same train-label ratios for direct budget comparison.

Dataset · task · splitATCNetDeepConvNetEEGConformerEEGNetShallowConvNet
1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%1%2%5%10%25%50%100%
HGDLeft/Right/Hand/Foot/TongueCS-0.23-0.060.1910.2320.360.5030.6220.2520.1110.1120.120.4920.5820.5730.000.020.080.3710.5310.6310.6910.1330.0330.0830.090.180.430.500.2810.1120.020.2130.4330.470.43
Meng2019Cursor Control + RestCS0.0720.1330.1230.2010.2210.2620.2920.040.070.000.010.030.080.110.0730.1910.1810.1930.170.220.283-0.070.120.010.090.2130.2330.250.1010.1420.1420.2020.2120.3010.351
OpenBMILeft/Right Hand MICS0.5220.5620.6030.6320.6810.7010.7320.470.500.6410.6510.6820.6820.7230.6910.6810.6120.6230.650.680.710.5030.500.500.610.6630.6830.710.470.5530.510.480.520.660.731
PhysioNet-MIHand vs Feet MICS-0.0130.00T-0.040.0430.050.2110.2730.00T-0.050.010.0720.2110.2120.301-0.250.1610.171-0.010.1720.2030.270.00T0.00T0.0320.020.0730.130.09-0.03-0.060.0230.1110.070.180.292
EEGMatMental Arithmetic vs RestCS0.400.420.500.5210.5010.500.500.5030.5410.5810.400.480.480.500.7010.420.500.5030.5030.6610.6520.5030.50T0.500.5030.5030.5030.500.5030.50T0.500.5030.5030.6320.691
TUSLBackground Slowing DetectionCS0.50T0.50T0.500.500.500.500.500.50T0.330.66T0.5630.7010.54T0.6310.000.50T0.360.6220.5030.54T0.5820.50T0.420.66T0.400.430.5230.5130.50T0.6710.500.6310.5520.480.48
ERP-BCIP300 Target DetectionCS-------0.0210.0510.0130.0030.003-0.000.00-0.020.0220.0120.0510.000.0030.000.00T0.000.000.0020.0810.1210.1210.00T0.0030.071-0.020.0120.0220.022
Kaggle-INRIAP300 Feedback CorrectnessCS-------0.080.390.6120.6710.5410.6010.5910.50T0.50T0.500.5030.5030.5220.5730.4230.4130.5530.500.480.460.5820.50T0.50T0.6410.5220.5020.4730.48
SEEDEmotion RecognitionCS-0.19-0.0130.0730.1320.2910.3030.2930.02-0.050.0820.1130.120.180.200.033-0.070.030.100.240.3210.3020.0420.052-0.02-0.040.2830.270.270.0510.1110.1310.2810.2920.3120.321