Bridging Self-Supervised Representation Learning and Imaging Inverse Problems

A spectral and structural analysis quantifying the connections between SSRL methods and self-supervised imaging losses

0.993
Spectral Correlation
47.0%
Physics MSE Reduction
0.651
Subspace Alignment
8
Methods Compared

Invariance Spectrum Comparison

Comparing eigenspectra of SSRL augmentation covariance and measurement operator.

Normalized Eigenspectra

Key Metrics

Spectral Correlation

0.993

Extremely high correlation between SSRL and measurement spectra

SSRL Effective Rank

64

All dimensions perturbed by augmentations

Measurement Effective Rank

32

Matches number of measurements m

Masking Principle Unification

MAE-style masking vs physics-aware reconstruction across masking ratios.

Reconstruction MSE vs Masking Ratio

Physics Advantage by Ratio

Reconstruction Convergence

Comparing convergence speed with zero vs SSRL-mean initialization.

Measurement Loss During Reconstruction

Transfer Performance

Reconstruction MSE across noise levels for different initialization strategies.

MSE vs Noise Level

Method Taxonomy

Characterizing 8 methods along invariance, masking, equivariance, and physics axes.

Method Feature Profiles

MethodTypeInvarianceMaskingEquivariancePhysics
SimCLRSSRL1.000.000.300.00
BYOLSSRL1.000.000.500.00
DINOSSRL0.800.200.600.00
MAESSRL0.201.000.100.00
Noise2SelfImaging0.300.800.200.70
EIImaging0.100.001.001.00
SSDUImaging0.200.500.300.90
Noisier2NoiseImaging0.100.000.200.80

Key Findings

Spectral Correspondence

SSRL augmentation and measurement operator spectra are highly correlated (0.993), suggesting both exploit similar geometric properties of signal spaces.

Masking Duality

Physics-aware reconstruction achieves 47% lower MSE than MAE-style mean-fill at 50% masking, but both share the same foundation of partial observation.

Transfer Potential

SSRL-initialized reconstruction provides comparable or improved convergence, with mean subspace alignment of 0.651 between domains.

Unified Taxonomy

Four axes (I, M, E, P) characterize all 8 methods, with equivariance serving as the primary bridge between SSRL and imaging.