Learned Regularization Methods

• Learned regularization is a framework where inductive biases are inferred from data or meta-training, enabling flexible, task-specific priors for inverse problems.
• The approach encompasses both explicit penalties, such as learned Gaussian priors for CNN kernels, and implicit methods like plug-and-play denoisers in iterative solvers.
• Empirical results show improved reconstruction accuracy and efficiency across domains like medical imaging and radio interferometry, while addressing challenges such as distribution mismatch and stability.

Learned regularization denotes a class of methods in which the regularizing bias is inferred from data, prior models, or meta-training rather than specified solely by a fixed analytic penalty. In the literature represented here, it appears as learned Gaussian priors over weights or convolution kernels, learned deformation priors, learned denoisers inserted into optimization algorithms, learned proximal operators in unrolled solvers, function-space regularization induced by architectural design, sampling distributions that implicitly define Bayes estimators, and meta-learned optimizer dynamics that internalize the effects of explicit regularizers (Feinman et al., 2019, Reithmeir et al., 2024, Terris et al., 2022, Babyale et al., 24 May 2026, Sahoo et al., 10 Oct 2025). Its unifying role is to narrow the feasible solution set toward structures that are plausible for a task or domain, while retaining enough flexibility to accommodate data fidelity and transfer across instances.

1. Conceptual scope

In classical inverse problems and registration, regularization is usually written as

$\min_x f(x) + \lambda R(x)$

or, in deformation form,

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$

with $R$ chosen analytically, for example as Tikhonov smoothness, total variation, bending energy, or volume penalties (Reithmeir et al., 2024). Learned regularization replaces or augments this analytic term by a prior inferred from data. In the most explicit formulation, one writes

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$

or uses a variational objective whose KL term measures deviation from a learned deformation prior (Reithmeir et al., 2024).

A second, equally important formulation is implicit. In plug-and-play and related methods, the regularizer is not given by a closed-form functional at all; a trained denoiser or proximal-like map is inserted into an iterative algorithm and acts as the prior through its action on iterates. AIRI in radio interferometry is a canonical example: it replaces a handcrafted proximal operator by a denoiser $D_\sigma$ inside forward–backward splitting, while retaining a data-fidelity gradient step driven by the measurement operator (Terris et al., 2022). The spectral-model analysis of learned inverse methods makes the same point in another language: many methods can be interpreted as learning spectral filters whose bias and stability depend on the training distributions for signal and noise (Burger et al., 2023).

The notion is broader still. In reinforcement learning, learned Fourier features do not regularize weights directly; they alter the Neural Tangent Kernel and therefore the induced RKHS norm, yielding frequency-selective functional regularization controlled by the initialization variance $\sigma^2$ of the learned basis (Li et al., 2021). In synthetic-data inverse problems, the sampling law itself becomes the regularizer: minimizing mean-squared empirical risk makes the learned inverse converge to the conditional expectation $\mathbb{E}_\mu[m\mid d]$, so the choice of sampling measure $\mu$ defines the implicit regularization operator (Babyale et al., 24 May 2026). This suggests that “learned regularization” is best understood as a family of mechanisms for encoding data-derived inductive bias, rather than a single algorithmic pattern.

2. Explicit learned priors and penalty constructions

One major line of work learns a regularizer as an explicit quadratic or probabilistic penalty. In “Learning a smooth kernel regularizer for convolutional neural networks” (Feinman et al., 2019), each vectorized convolution kernel $w=\mathrm{vec}(W)$ is given a learned Gaussian prior

$p(w)=\mathcal{N}(0,\Sigma),$

which induces the quadratic penalty

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$0

The full loss becomes

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$1

with layer-specific covariances $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$2 estimated by empirical Bayes from kernels of previously trained CNNs (Feinman et al., 2019). Standard $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$3 weight decay is recovered only in the isotropic special case $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$4; the learned, non-diagonal $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$5 instead encodes spatial correlations and smoothness.

A closely related construction appears in compressed sensing with Deep Image Prior, where the learned regularizer is a Gaussian prior over network weights rather than filters:

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$6

In practice the paper uses a layerwise diagonal approximation, with shared mean and variance per layer,

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$7

and augments the DIP objective by both total variation and this learned term (Veen et al., 2018). Here the prior is estimated from optimized weights obtained from measurements of similar images, without requiring ground-truth images for prior estimation (Veen et al., 2018).

Medical image registration extends the same principle from weights to deformations. The 2024 review describes learned deformation priors $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$8, variational approximations $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda R(\phi),$$9, PCA-based and autoencoder-based deformation manifolds, adversarial plausibility losses of the form

$R$0

and learned operators or metrics in LDDMM-like models (Reithmeir et al., 2024). In that setting, learned regularization may act explicitly as $R$1, implicitly via diffeomorphic architectures and SVF parameterizations, or conditionally through hypernetworks and conditional normalization layers that adapt regularization weights at test time (Reithmeir et al., 2024).

A more structured explicit energy appears in electrocardiographic imaging, where regularization couples space and time. The learned term is a temporal Fields-of-Experts prior with learned kernels $R$2, learned potentials $R$3, and an FEM discretization on unstructured cardiac meshes:

$R$4

This combines spatial regularity with a learned temporal prior over activation dynamics, rather than relying only on spatial smoothing (Haas et al., 7 Feb 2026).

3. Implicit priors, plug-and-play operators, and unrolled regularization

A second major paradigm replaces an explicit $R$5 by a learned operator inside an optimization algorithm. AIRI in radio-interferometric imaging formulates the data model as $R$6 and alternates

$R$7

with $R$8 and a learned denoiser $R$9 trained at a noise level matched to the target dynamic range (Terris et al., 2022). Convergence is tied to firm nonexpansiveness of the denoiser, which AIRI enforces approximately through a Jacobian spectral-norm penalty during training (Terris et al., 2022).

Planet cartography uses a directly analogous unrolled structure. Spin–orbit tomography is written as $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$0, and each ISTA-like stage performs a gradient step followed by a learned proximal map $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$1:

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$2

The learned proximal is trained on procedurally generated Earth-like surfaces and clouds, so the regularizing effect is the projection of iterates onto a learned manifold of plausible albedo maps (Ramos et al., 2020).

Microwave tomography provides a diffusion-model variant. SSD-Reg combines a physics-based data-consistency term with a single-step diffusion regularizer

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$3

where $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$4 is the pretrained diffusion noise predictor and $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$5 (Tong et al., 11 Aug 2025). The result is a PnP-style variational scheme driven by exact Fréchet derivatives of the electromagnetic forward operator and a one-step learned prior gradient, rather than full reverse diffusion sampling (Tong et al., 11 Aug 2025).

NETT, or Network Tikhonov Regularization, occupies an intermediate position. Its objective remains variational,

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$6

but the regularizer is defined by a trained network, for example

$$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$7

The discretized NETT analysis explicitly tracks discretization of the data space, solution space, forward operator, and network, and proves asymptotic convergence for decreasing noise and discretization errors (Antholzer et al., 2020).

Plug-and-play learned regularization also appears outside reconstruction. In curvilinear-structure segmentation, a residual U-Net reconnecting operator $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$8 replaces the proximal map of a box constraint after the iterate becomes “almost binary,” within a forward–backward primal–dual scheme (Carneiro-Esteves et al., 2024). The regularizer is learned not from annotated target images but from synthetic connected/disconnected pairs, so the prior specifically encodes connectivity preservation rather than generic denoising (Carneiro-Esteves et al., 2024).

4. What is learned: priors, spectra, hyperparameters, and optimization dynamics

The object being learned varies substantially across the literature. In some cases it is a covariance, precision, or quadratic form, as in SK-reg and learned weight priors (Feinman et al., 2019, Veen et al., 2018). In some it is a denoiser or proximal surrogate, as in AIRI, planet cartography, microwave tomography, and reconnecting segmentation (Terris et al., 2022, Ramos et al., 2020, Tong et al., 11 Aug 2025, Carneiro-Esteves et al., 2024). In registration it may be a deformation prior, a spatially varying regularization weight $$\min_{\phi} D(I_f, I_m \circ \phi) + \lambda \big(-\log p_\theta(\phi)\big),$$9, a learned RKHS metric, or a diffeomorphic parameterization (Reithmeir et al., 2024).

A spectral perspective clarifies this diversity. The 2023 spectral-model chapter studies learned inverse methods as coordinate-wise filters

$D_\sigma$0

for the singular system of the forward operator. For supervised MSE learning, the optimal learned filter is

$D_\sigma$1

where $D_\sigma$2 is prior energy per mode and $D_\sigma$3 is noise energy per mode (Burger et al., 2023). In this view, learned regularization amounts to learning a distribution-dependent Tikhonov filter.

The sampling-distribution analysis pushes this idea further. If synthetic training pairs are generated by sampling $D_\sigma$4 and $D_\sigma$5, then minimizing squared empirical risk converges, in the infinite-data limit, to

$D_\sigma$6

The measure $D_\sigma$7 is therefore an implicit regularizer, and the paper shows that Gaussian, Laplace, TV-increment, and uniform sampling induce biases analogous to $D_\sigma$8, $D_\sigma$9, TV, or no explicit structural regularization, respectively (Babyale et al., 24 May 2026).

Two additional generalizations are noteworthy. First, in quantitative pulse-echo speed-of-sound imaging, the learned object is a linear operator $\sigma^2$0 chosen in closed form to minimize average reconstruction error over a simulated distribution of anatomies,

$\sigma^2$1

This can be interpreted as learning the entire inverse map, or equivalently a learned Tikhonov operator (Yolgunlu et al., 2024). Second, in learned optimizers, regularization is encoded in the optimizer dynamics themselves. The outer objective includes a smoothness term on the update map and a meta-regularizer inspired by SAM, GSAM, or GAM,

$\sigma^2$2

so that the learned optimizer later seeks minima with regularization-like geometry even when no explicit regularizer is applied to the optimizee loss at meta-test (Sahoo et al., 10 Oct 2025).

5. Empirical evidence across domains

The empirical record in these papers is heterogeneous but consistently aimed at settings where handcrafted regularizers are too rigid or too weak. In small-data visual recognition, the smooth kernel regularizer improves a 10-class silhouettes task from $\sigma^2$3 accuracy under $\sigma^2$4 to $\sigma^2$5 under SK-reg, corresponding to an approximately 55% relative improvement, and improves Tiny ImageNet accuracy from $\sigma^2$6 to $\sigma^2$7 when transferring a kernel prior learned from silhouettes (Feinman et al., 2019).

In radio interferometry, AIRI-$\sigma^2$8 is reported as about 3 dB above uSARA in SNR and about 1–2 dB above AIRI-$\sigma^2$9 in logSNR across tested durations, while GPU AIRI is 5–10× faster than uSARA (Terris et al., 2022). In microwave tomography, SSD-Reg converges in about 200 iterations, has average runtime 45.22 s versus 408.72 s for INR+TV, and preserves strong SSIM and PSNR even at 30% Gaussian noise (Tong et al., 11 Aug 2025). In quantitative pulse-echo speed-of-sound imaging, learned regularization reduces liver speed-of-sound bias standard deviation from about 20–24 m/s for gradient regularization to about 5–6 m/s in the full-wave setting, and to 1.1 m/s in the strictly linear setting (Yolgunlu et al., 2024).

Plug-and-play reconnecting regularization targets topology rather than region overlap. On vascular segmentation, it reduces connected-component error by approximately 90% in 2D and 70% in 3D relative to variational baselines, while also generalizing to road cracks and porcine corneal cells (Carneiro-Esteves et al., 2024). In building-mask polygonization, the learned adversarial regularizer keeps IoU and accuracy close to strong FCN baselines while producing more rectilinear, polygon-ready footprints and improving Mask R-CNN outputs when used as post-regularization (Zorzi et al., 2020).

Learned regularization also improves robustness in inverse problems without large paired datasets. In CS-DIP, a learned Gaussian prior over network weights yields percent MSE reductions on chest X-rays that grow with noise, reaching 37.4% at $\mathbb{E}_\mu[m\mid d]$0 and $\mathbb{E}_\mu[m\mid d]$1 relative to no learned regularization (Veen et al., 2018). In diffuse optical tomography, Learned-SVD preserves absorption contrast and achieves higher TPR than Elastic Net or Bregman under increasing Gaussian noise, while requiring no hand-tuned regularization parameter at inference (Benfenati et al., 2021). In reinforcement learning, learned Fourier features outperform MLP baselines in 6 of 8 state-based SAC tasks and give dramatic sample-efficiency gains in pixel-based SAC+RAD, especially through stabilization of value learning (Li et al., 2021).

6. Limitations, controversies, and directions

A persistent limitation is distribution mismatch. The sampling-distribution analysis makes this especially explicit: a mismatched sampling law degrades reconstruction quality “in ways that neither more expressive architectures nor augmented physics residuals can fully correct” (Babyale et al., 24 May 2026). The same issue appears in multiple domains: SK-reg may over-smooth tasks dominated by high-frequency textures (Feinman et al., 2019); exoplanet mapping can hallucinate Earth-like morphologies when the learned prior is misaligned with the true planet class (Ramos et al., 2020); medical registration priors may overfit specific anatomies or simulation designs (Reithmeir et al., 2024); and SSD-Reg can under-regularize semantically complex anatomy because its diffusion prior was trained on generic polygons and Bézier shapes rather than anatomy-specific data (Tong et al., 11 Aug 2025).

A second limitation is theoretical. In AIRI, the denoiser acts as an implicit prior and need not correspond to an explicit $\mathbb{E}_\mu[m\mid d]$2 (Terris et al., 2022). In curvilinear PnP segmentation, convergence of the learned reconnecting operator is not proved, and the authors identify learning a maximally monotone operator as future work (Carneiro-Esteves et al., 2024). Learned reconstruction algorithms can also introduce instabilities that are distinct from the ill-posedness of the forward operator itself. The analysis of learned post-processing and learned unrolling shows that hallucinations and discontinuities can arise from the learned component, and proposes mixing coefficients $\mathbb{E}_\mu[m\mid d]$3 to preserve continuity and convergent regularization behavior (Nayak, 2021).

A third debate concerns what must be modeled explicitly rather than learned implicitly. The 2025 noise-model analysis shows that for colored noise, learned variational schemes without the correct noise weighting can exhibit a provable performance gap relative to the optimal affine reconstruction; weighted Tikhonov, Lavrentiev, and quadratic schemes recover the optimum only when the noise covariance is modeled or co-learned (Banert et al., 14 Oct 2025). The 2026 remark on classical and learned Tikhonov similarly emphasizes that parameter misspecification across noise levels may be milder than often assumed, but infinite-dimensional learned $\mathbb{E}_\mu[m\mid d]$4 regularization with non-injective transforms can be pathological unless one works in a suitably discretized finite-dimensional setting (Behrens et al., 1 Apr 2026).

Current directions therefore combine stronger theory with broader priors. The literature here points to richer generative priors and structured precision models for CNN kernels (Feinman et al., 2019), foundation-model and physics-informed regularization for registration (Reithmeir et al., 2024), explicit discretization-aware theory for neural regularizers in NETT (Antholzer et al., 2020), adaptive or co-learned noise models (Banert et al., 14 Oct 2025), and meta-learned optimizers that encode flatness-seeking behavior without explicit test-time regularization (Sahoo et al., 10 Oct 2025). A plausible implication is that learned regularization is evolving from a narrow replacement for handcrafted penalties into a general framework for learning the geometry, statistics, and algorithmic action of priors themselves.