Deep learning has revolutionized our ability to solve inverse problems—such as reconstructing a crisp image from a noisy sensor or filling in missing data points. Traditionally, training these models requires a supervised paradigm: a large dataset pairing corrupted or incomplete measurements with their corresponding "clean" ground-truth references.
However, in many real-world scenarios, ground truth is a luxury we cannot afford. In cardiac MRI, rapid organ movement and signal decay make fully sampled scans physically impossible to acquire. In low-light microscopy, capturing a noise-free reference requires exposing delicate biological samples to damaging levels of radiation. As noted by Shanshan Wang et al. in LANTERN: Learn analysis transform network for dynamic magnetic resonance imaging, breaking this dependency on perfect reference images is critical to unlocking vast amounts of clinical and scientific data that currently cannot be used for supervised training.
To solve this, a powerful paradigm has emerged: self-supervised measurement splitting (or data partitioning). This technique allows neural networks to learn high-fidelity signal reconstruction directly from incomplete, noisy measurements, without ever seeing a single clean example.
The Core Mechanism: Splitting the Incomplete
If we only have access to an incomplete measurement vector $y$, we cannot simply train a network to map $y \to y$. Doing so would cause the network to learn a trivial identity function, memorizing the noise or the missing gaps instead of learning how to reconstruct the underlying signal.
Measurement splitting bypasses this failure mode by partitioning the available measurements into two or more disjoint subsets:
$$\Omega \cap \Lambda = \emptyset$$
- The Input Subset ($y_\Omega$): This portion of the data is fed into the reconstruction network to generate an estimate of the full signal.
- The Target Subset ($y_\Lambda$): This portion is withheld from the network during the forward pass. It is used exclusively to calculate the training loss.
[ Incomplete Measurements (y) ]
│
┌───────┴───────┐
▼ ▼
Input (y_Ω) Target (y_Λ)
│ │
▼ │
[ Network ] │
│ │
▼ ▼
[ Reconstructed ] ──► [ Loss Calculation ]
During training, the network takes $y_\Omega$ and attempts to reconstruct the full signal. The resulting output is then projected back into the measurement domain, and the loss function evaluates how well the network's prediction matches the held-out target data $y_\Lambda$.
As pioneered by Burhaneddin Yaman et al. in Self-Supervised Physics-Based Deep Learning MRI Reconstruction Without Fully-Sampled Data, this simple act of withholding data forces the network to generalize. Because the network cannot see $y_\Lambda$ in the input, it cannot rely on a trivial identity mapping; it must learn to interpolate the missing information.
Why It Works: The Intuition of Structural Correlation
Why does minimizing a loss on held-out, incomplete measurements yield a clean, high-fidelity reconstruction? The answer lies in the fundamental difference between the structure of real-world signals and the structure of noise or sampling artifacts.
1. Signal Redundancy vs. Artifact Independence
Physical signals—whether they are anatomical structures in an MRI scan, pixels in a natural image, or trends in a time series—possess immense internal correlation. Nearby frequencies, pixels, or time steps are highly dependent on one another, governed by underlying physical laws or spatial smoothness.
Conversely, measurement noise and undersampling artifacts are typically uncorrelated or follow distinct geometric patterns that do not align with the true signal. When the network is forced to predict the held-out measurements $y_\Lambda$ from the input measurements $y_\Omega$, the only consistent strategy that minimizes the loss across the entire dataset is to learn the underlying, shared structural regularities of the signal. The network cannot predict the random noise or the specific artifact patterns of $y_\Lambda$ from $y_\Omega$, so it learns to ignore them, acting as an optimal interpolator.
2. Theoretical Guarantees and the "Data Tax"
This intuition is backed by rigorous theory. Victor Sechaud et al. demonstrated in Equivariant Splitting: Self-supervised learning from incomplete data that if the measurement process is sufficiently diverse, partitioning a single set of incomplete observations is mathematically sufficient to recover the true underlying signal.
However, learning without ground truth does incur a "data tax." In Analyzing the Sample Complexity of Self-Supervised Image Reconstruction Methods, Tobit Klug et al. showed that while self-supervised partitioning requires more training samples to reach the same initial accuracy as supervised models, this performance gap completely vanishes as the training set size grows.
Variations Across Domains
The philosophy of measurement splitting is highly versatile and has been adapted across various data modalities by modifying how the partition is constructed.
Accelerated MRI and Compressive Imaging
In accelerated MRI, data is acquired in the frequency domain (k-space). Self-supervised models partition these frequency coordinates into disjoint sets. To maximize performance, the authors of Dual-Domain Self-supervised Learning and Model Adaption for Deep Compressive Imaging (Yuhui Quan et al.) introduced a dual-domain framework. It enforces consistency not just on the partitioned frequency measurements, but also applies regularization in the transformed image domain, matching the fidelity of fully supervised alternatives.
Remarkably, this partitioning can even happen within a single acquisition. In Zero-Shot Self-Supervised Learning for MRI Reconstruction, Burhaneddin Yaman et al. showed that by partitioning a single patient's undersampled data into internal training, loss, and validation subsets, a network can be optimized on the fly for that specific scan, eliminating the need for external training databases entirely.
Photon-Driven Denoising
The splitting concept extends beautifully to discrete event data. In Image Denoising and the Generative Accumulation of Photons, Alexander Krull et al. split the discrete photon arrivals of a single low-light observation into input and target subsets using a binomial distribution. They proved that training a model to predict where the target photons will land based on the input photons is mathematically equivalent to training an optimal minimum mean-squared error (MMSE) denoiser.
Tabular and Time-Series Imputation
Beyond physical imaging, the same masking-as-partitioning strategy solves data gaps in tabular and continuous data. Yusuke Tashiro et al. developed CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation, which randomly masks observed time-series values to act as artificial targets for a diffusion model. Similarly, in ReMasker: Imputing Tabular Data with Masked Autoencoding, Tianyu Du et al. hide subsets of known tabular entries, creating a self-contained supervised task that forces the network to learn the global correlations inherent to the table.
Overcoming the Splitting Penalty
While traditional measurement splitting is highly effective, it suffers from a fundamental trade-off: the splitting penalty. By withholding $y_\Lambda$ from the input network, the model is forced to reconstruct the signal from a strictly smaller, more degraded subset of data than what was actually collected by the sensor. This can lead to unstable training or a loss of fine, low-contrast structural details.
To resolve this limitation, recent work has moved toward eliminating rigid partitions. In AeSPa : Attention-guided Self-supervised Parallel Imaging for MRI Reconstruction, Jinho Joo et al. proposed a framework that feeds all acquired measurements into the network as both the input and the target.
To prevent the network from simply learning the identity mapping, they replace physical data splitting with iterative masking and selective attention mechanisms inside the network architecture. By dynamically blinding the network to specific elements during the attention calculation, the model can exploit the full signal capacity of the collected data without the risk of trivial memorization, representing the next evolution in ground-truth-free reconstruction.