# Proximal-Mismatch Fine-Tuning

- ID: 938
- Canonical URL: https://synthcore.org/idea/938/proximal-mismatch-fine-tuning
- API JSON: https://synthcore.org/api/idea/938.json
- API Markdown: https://synthcore.org/api/idea/938.md
- Verification status: failed_benchmark
- Source: [arXiv:2607.14894](https://arxiv.org/abs/2607.14894)
- Category: regularization
- Solves: accuracy, stability, sample-efficiency
- ML areas: loss, fine-tuning, regularization, optimizer
- Math tags: convex-analysis, optimization, numerical-analysis
- Ratings: usefulness 7/10; difficulty 5/10; novelty 6/10

## Idea description

Fine-tune a denoiser by matching its action to a target-domain proximal operator, instead of minimizing only pixelwise denoising error. Apply the loss on the intermediate states and noise levels actually encountered by the downstream iterative solver, so the adaptation directly reduces the error that controls PnP reconstruction stability.

## Mathematical statement

The paper formulates the regularized target problem as $F(\boldsymbol{x})=\eta f(\boldsymbol{x})+R(\boldsymbol{x})$, where $f$ is the data-fidelity function, $R$ is the target-domain regularizer, and $\eta>0$ weights data consistency. The ideal denoising operator is $\mathsf{D}_{\star}=\operatorname{prox}_{R_{\star}}$, with $\operatorname{prox}_{\gamma R_{\star}}(\boldsymbol{v})=\arg\min_{\boldsymbol{z}}\{R_{\star}(\boldsymbol{z})+(2\gamma)^{-1}\|\boldsymbol{z}-\boldsymbol{v}\|_2^2\}$; $\gamma>0$ is the proximal step size. The deployed network is $\widehat{\mathsf{D}}$. Define the proximal mismatch at input $\boldsymbol{v}$ as $\delta(\boldsymbol{v})=\|\widehat{\mathsf{D}}(\boldsymbol{v})-\mathsf{D}_{\star}(\boldsymbol{v})\|_2$. The paper's key transfer principle is that PnP stationarity has an $\mathcal{O}(1/K)$ optimization term plus an additive mismatch term proportional to the average squared error: $\min_{0\le k<K}\operatorname{Stat}(\boldsymbol{x}_k)\lesssim C/K+(1/K)\sum_{k=0}^{K-1}\delta(\boldsymbol{v}_k)^2$, where $K$ is the number of PnP iterations, $C$ collects initial-objective and step-size factors, and $\boldsymbol{v}_k$ is the denoiser input. We use this as the training objective: minimize empirical $\delta^2$ over target-domain solver states.

## Key formulas

- $$F(\boldsymbol{x})=\eta f(\boldsymbol{x})+R(\boldsymbol{x})$$
- $$\mathsf{D}_{\star}=\operatorname{prox}_{R_{\star}},\qquad \operatorname{prox}_{\gamma R_{\star}}(\boldsymbol{v})=\arg\min_{\boldsymbol{z}}\left\{R_{\star}(\boldsymbol{z})+\frac{1}{2\gamma}\|\boldsymbol{z}-\boldsymbol{v}\|_2^2\right\}$$
- $$\delta(\boldsymbol{v})=\left\|\widehat{\mathsf{D}}(\boldsymbol{v})-\mathsf{D}_{\star}(\boldsymbol{v})\right\|_2$$
- $$\min_{0\le k<K}\operatorname{Stat}(\boldsymbol{x}_k)\;\lesssim\;\frac{C}{K}+\frac{1}{K}\sum_{k=0}^{K-1}\left\|\widehat{\mathsf{D}}(\boldsymbol{v}_k)-\mathsf{D}_{\star}(\boldsymbol{v}_k)\right\|_2^2$$

## Implementation notes

(1) Integration point: use an image denoiser $D_\theta$ inside a PnP proximal-gradient reconstruction loop. For measurements $y$, forward operator $A$, and differentiable data loss $f(x)=\tfrac12\|Ax-y\|_2^2$, form $v_k=x_k-\gamma\eta A^T(Ax_k-y)$ and update $x_{k+1}=D_\theta(v_k,\sigma_k)$. During adaptation, modify the denoiser training loss, not the PnP optimizer.

(2) Pseudocode: sample a target-domain clean image $x$, generate a measurement $y=Ax+e$, and initialize $z_0$. For $k=0,\ldots,K-1$, compute $v_k=z_k-\gamma\eta A^T(Az_k-y)$, compute $u_k=D_\theta(v_k,\sigma_k)$, and compute a target proximal output $t_k$. If a differentiable target proximal teacher $D_\star$ is available, set $t_k=D_\star(v_k,\sigma_k)$. Otherwise obtain $t_k$ from a strong frozen teacher denoiser, a converged domain-specific PnP solver, or an inner optimization $t_k\approx\arg\min_z\{R_\star(z)+(2\gamma)^{-1}\|z-v_k\|^2\}$. Optimize $L_{prox}=\frac1{BK}\sum_{b,k}w_k\|D_\theta(v_{b,k},\sigma_k)-t_{b,k}\|_2^2+\lambda L_{MSE}$, then update $z_{k+1}=\operatorname{stopgrad}(u_k)$ while collecting states. Use $w_k$ to emphasize late or high-mismatch states.

(3) Computed from the paper's mathematics: the training signal is squared proximal mismatch and the predicted reconstruction floor is its empirical average. Estimated empirically: $D_\star$ or its proxy, the state distribution $v_k$, and $\operatorname{Stat}(x_k)$, measured using the data-gradient norm or reconstruction-objective decrease. Compare proximal matching against ordinary clean-target MSE and a mixed loss.

(4) First experiment: adapt a small DnCNN or U-Net denoiser with 1-10 target-domain images for Gaussian deblurring on BSD68 or a shifted medical or satellite subset. Run 20-step PnP-PGD with identical $A$, noise, step sizes, and parameter count. Test ordinary MSE adaptation, proximal matching, and a 50/50 mixture. Success means higher final reconstruction PSNR and SSIM, faster decrease of $F(x_k)$, and a lower measured stationarity or data-consistency residual at equal PnP iterations, especially in the few-shot regime.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: no_effect
- Verdict: Built and ran a reproducible scalar PnP proximal-mismatch experiment in experiment.py. All three quantitative predictions were confirmed: proximal coefficient error was 3.3e-16, mismatch scaling had log-log slope 2.000 versus predicted 2, zero-regularization mismatch was 0, and the stability transition bracketed the predicted boundary a=2.0 between 1.99 and 2.01. Proximal matching also beat clean-MSE in this exactly solvable quadratic toy, but this does not establish real-image performance.

### Mechanism check

- Verdict: Built and ran a reproducible scalar PnP proximal-mismatch experiment in experiment.py. All three quantitative predictions were confirmed: proximal coefficient error was 3.3e-16, mismatch scaling had log-log slope 2.000 versus predicted 2, zero-regularization mismatch was 0, and the stability transition bracketed the predicted boundary a=2.0 between 1.99 and 2.01. Proximal matching also beat clean-MSE in this exactly solvable quadratic toy, but this does not establish real-image performance.
- Confidence: 9/10
- Limitations: No image denoiser, real BSD/medical data, learned nonlinear network, approximate teacher, noisy inverse operator, SSIM, GPU training, or few-shot generalization was tested. The proximal teacher was analytically known, so the practical gain is evidence only for the isolated quadratic mechanism.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.8863
- Paired wins: 5/8
- Benchmark verdict: no measurable effect

## Artifacts

- [bench_experiment.py](https://synthcore.org/code/397/bench_experiment.py)
- [bench_report.json](https://synthcore.org/code/397/bench_report.json)
- [experiment.py](https://synthcore.org/code/397/experiment.py)
- [report.md](https://synthcore.org/code/397/report.md)
- [report_bench_2026-08-31T111048.md](https://synthcore.org/code/397/report_bench_2026-08-31T111048.md)
- [results.json](https://synthcore.org/code/397/results.json)
- [Download all files as ZIP](https://synthcore.org/download/397)

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.