A spectral and structural analysis quantifying the connections between SSRL methods and self-supervised imaging losses
Comparing eigenspectra of SSRL augmentation covariance and measurement operator.
Extremely high correlation between SSRL and measurement spectra
All dimensions perturbed by augmentations
Matches number of measurements m
MAE-style masking vs physics-aware reconstruction across masking ratios.
Comparing convergence speed with zero vs SSRL-mean initialization.
Reconstruction MSE across noise levels for different initialization strategies.
Characterizing 8 methods along invariance, masking, equivariance, and physics axes.
| Method | Type | Invariance | Masking | Equivariance | Physics |
|---|---|---|---|---|---|
| SimCLR | SSRL | 1.00 | 0.00 | 0.30 | 0.00 |
| BYOL | SSRL | 1.00 | 0.00 | 0.50 | 0.00 |
| DINO | SSRL | 0.80 | 0.20 | 0.60 | 0.00 |
| MAE | SSRL | 0.20 | 1.00 | 0.10 | 0.00 |
| Noise2Self | Imaging | 0.30 | 0.80 | 0.20 | 0.70 |
| EI | Imaging | 0.10 | 0.00 | 1.00 | 1.00 |
| SSDU | Imaging | 0.20 | 0.50 | 0.30 | 0.90 |
| Noisier2Noise | Imaging | 0.10 | 0.00 | 0.20 | 0.80 |
SSRL augmentation and measurement operator spectra are highly correlated (0.993), suggesting both exploit similar geometric properties of signal spaces.
Physics-aware reconstruction achieves 47% lower MSE than MAE-style mean-fill at 50% masking, but both share the same foundation of partial observation.
SSRL-initialized reconstruction provides comparable or improved convergence, with mean subspace alignment of 0.651 between domains.
Four axes (I, M, E, P) characterize all 8 methods, with equivariance serving as the primary bridge between SSRL and imaging.