{"id":"9aa02919-bbd3-4e32-9e9d-b2e77337ac35","arxiv_id":"2506.07008","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-step trained neural network predicts regularization parameters for noisy-operator inverse scattering, yielding faster linear sampling method reconstructions with contrast at or above manually tuned Morozov regularization in synthetic tests.","lead":"This paper trains a small neural network to choose regularization parameters for linear sampling method imaging when the measurement operator is noisy, replacing slow manual tuning. In synthetic crack-imaging tests, the network produces images with equal or better contrast than manually tuned Morozov regularization, and could speed up real-time ultrasonic imaging.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Step 2 enhancement claim rests on a stopping-time artifact: Eq. (22) has no interior minimum in α and is minimized at α→0, so the reported contrast gains are selected by Algorithm 2's thresholds, not by the stated objective.","rationale":"I read the paper's central claim in good faith: a two-step learned regularization scheme in which Step 1 learns coarse Morozov maps and Step 2 refines them by minimizing the Tikhonov loss, with synthetic evidence that the informed variant improves LSM image contrast beyond manually tuned Morozov regularization. The load-bearing condition is that Step 2 genuinely optimizes something that tracks image quality. Mathematically, this condition fails: the Tikhonov loss evaluated at its own minimizer is monotone in α and vanishes as α→0, so minimizing it pushes regularization parameters to zero. Whatever contrast improvement is observed must therefore come from the early-stopping rule, whose thresholds are not derived, tested for sensitivity, or tied to any independent measure of reconstruction fidelity. The reader identified exactly this weakness, and I agree with the conditional verdict: the empirical tables are interesting but the central claim is not established until this degeneracy is either removed by a better objective (e.g., one that includes a nontrivial penalty on ∥g∥^2 or uses a validation-based early stopping on a true quality metric) or shown to be robust by the concrete continuation test. I did not find a more load-bearing concern: the internal inconsistency between Eq. (15)/(32) and Eq. (19) is real but affects Step 1's loss-balancing derivation rather than the core enhancement claim, and the absence of error bars is a reporting weakness rather than the decisive mathematical gap. My recommendation is therefore to leave the reader's CONDITIONAL verdict unchanged while requiring the continuation experiment as a condition for acceptance.","tokens_in":26048,"tokens_out":7613,"duration_ms":92315,"concrete_test":"On the t3/25%-noise configuration reported in Tables 5 and 6, rerun the informed R-Net Step 2 with Algorithm 2's stopping criterion disabled and continue training for several multiples of N_st (e.g., 2x, 5x, 10x), logging the dense-grid contrast metrics C_mn and C_mx alongside the average predicted α_NN at each checkpoint. If contrast peaks only in a narrow epoch window around the original N_st and then deteriorates while J_img keeps decreasing toward machine precision, the reported enhancement is an early-stopping artifact rather than an objective-driven gain; if contrast remains high for all checkpoints past the trigger and α_NN stays bounded away from machine precision, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that discrepancy-informed R-Nets 'remarkably enhance image quality' is carried almost entirely by Step 2, whose objective is degenerate. For any fixed right-hand side u_L, with g_NN(α) as in Eq. (23), the Tikhonov loss in Eq. (22) evaluates to J_t(α) = Σ_j α |(u_j^*, u_L)|^2 / (α + d_j^2), where d_j are the singular values. This function is strictly increasing in α for α > 0 and tends to 0 as α → 0. Therefore the unconstrained minimization of Eq. (22) has no interior optimum: it drives α_NN toward 0, which amplifies the solution norm and does not correspond to any well-defined 'min-norm solution' producing contrast. The paper acknowledges that overfitting must be prevented and introduces Algorithm 2, but both J_img and the validation loss V_img in Eq. (24) share the same α→0 degeneracy, so the validation trajectory cannot detect the norm blow-up either. The reported improvements after Step 2 are thus a property of the particular stopping thresholds (σ_r = 5, σ_a = 10^-4, N_rms = 10000) and the chosen training duration, not a proven consequence of minimizing the imaging objective. No sensitivity analysis for these thresholds is given, and the central message of the paper is not independently supported once this dependence is exposed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-step deep-learning method for choosing per-right-hand-side Tikhonov parameters in linear sampling method (LSM) reconstructions when the scattering operator itself is noisy. In Step 1, an MLP is trained on low-resolution regularization maps produced by the Morozov discrepancy principle with a rough threshold η0; two modes are considered, a basic squared-error mode and a 'discrepancy-informed' mode that adds the discrepancy functional to the loss with adaptive weights. In Step 2, the network is fine-tuned by minimizing the Tikhonov imaging loss on the training right-hand sides, with an early-stopping rule based on training and validation loss trajectories. The method is tested on synthetic elastodynamic data for a plate with an evolving damage zone at 0%, 10%, and 25% noise, and the resulting LSM images are compared with manually tuned Morozov reconstructions through contrast metrics. The authors report that discrepancy-informed R-Nets generalize from 50×50 to 100×100 grids and that Step 2 increases contrast, especially at high noise.","tokens_in":26420,"tokens_out":10735,"duration_ms":101080,"significance":"The engineering problem is relevant: per-RHS Tikhonov regularization for LSM is a genuine bottleneck for real-time imaging, and the idea of learning a map from projected right-hand sides to α from coarse Morozov maps is attractive. The paper's strengths include complete algorithmic pseudocode, an explicit SVD-based architecture, a substantial synthetic testbed with five damage configurations and three noise levels, and an honest comparison with manually tuned Morozov baselines. If the Step-2 claim were supported, the practical contribution would be valuable. However, the current manuscript does not establish that Step 2 improves images as a consequence of the stated objective; the objective is degenerate in α, and the claimed gains appear to be selected by early-stopping thresholds. The discrepancy-functional equations also contain an internal inconsistency that affects the Step-1 labels and the loss-balancing derivation.","major_comments":[{"comment":"The Step 2 objective is degenerate in α and does not, as stated, drive the network toward higher-quality images. For a fixed right-hand side u_L, substituting (23) into (22) gives J_t(α)=Σ_j α |(u_j^*,u_L)|^2/(α+d_j^2), which is strictly increasing for α>0 and tends to 0 as α→0; the unconstrained minimization therefore pushes α_NN toward 0 and ||g_NN|| toward infinity, and no interior optimum exists. Because V_img in (24) is the same functional evaluated on a validation set, its trajectory cannot independently detect this norm blow-up. As written, the reported Step 2 contrast gains are selected by the particular thresholds σ_r=5, σ_a=10^{-4}, N_rms=10^4 and the learning rate 5×10^{-8}, not by a proven property of the objective. Please either add a penalty or constraint that gives (22) a nontrivial minimizer, or reframe Step 2 as an empirically early-stopped heuristic and provide a sensitivity analysis over these thresholds plus an independent quality measure on held-out configurations.","section":"3.2, Eqs. (22)–(24)"},{"comment":"The discrepancy functional is internally inconsistent. For the Tikhonov filter in (23), the Morozov condition ||F^δ g−u_L||=η||g|| is Σ_j (α²−η²D_jj²)/(α+D_jj²)² |(u_j^*,u_L)|²=0, but Eqs. (15) and (32) write the numerator as α−η²D_jj². The derivative in (19), 2D_jj²(α+η²)/(α+D_jj²)³, is the derivative of the α² version, not of the printed α version. This matters because (32) generates the Morozov labels used for Step 1 training and for the manually tuned baselines, while (19) is the basis of the proposed loss-balancing weights. The equations should be reconciled and the numerical results re-checked with the corrected Morozov functional.","section":"3.1.2 and 4.4, Eqs. (15), (19), (32)"},{"comment":"The evaluation of Step 2 is transductive: the network is trained on a subset of the RHS patterns of the same damage configuration, and the reported LSM images are then reconstructed from the full RHS dictionary of that same configuration, including the patterns used in Step 2. The claimed 'remarkable enhancement' therefore does not yet demonstrate generalization to unseen configurations. Please report results on a genuinely held-out configuration, or at least on RHS patterns excluded from both training steps, and state whether the contrast gains persist there.","section":"4.4, Figs. 18, 22, 26 and Tables 4–6"}],"minor_comments":[{"comment":"The summation upper limit is written as Ntrn, although the validation set is indexed by ν=1,…,Nν; this should be Nν.","section":"Eq. (24)"},{"comment":"The text states that the training runs for t∈[1,1000], while Step 2 is said to start at t=10001 and the caption of Fig. 17 says Step 1 ends at t=10^4; please reconcile the epoch counts.","section":"4.4, informed R-Net Step 1"},{"comment":"The statement that minimizing (22) 'minimizes the residual ... while maximizing the LSM indicator functional' is not a direct consequence of the displayed objective, since it trades residual against ||g||² and does not explicitly maximize 1/||g||; please rephrase.","section":"3.2, sentence after Eq. (23)"},{"comment":"The stopping rule uses OR conditions, so a flattening of either the training or the validation loss can stop training; a patience parameter or a more explicit combination of the conditions may better match the stated goal of detecting overfitting.","section":"Algorithm 2"},{"comment":"Reporting the time per epoch is useful, but the total wall-clock time until the stopping criterion is met, including the validation-loss evaluation in Step 2, would better support the real-time imaging claim.","section":"Table 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is positioned at the intersection of inverse problems and deep learning and fits the journal's scope. The main risk is the Step-2 claim: because the objective (22) is minimized at α→0, the reported contrast gains are not explained by the stated loss. I would ask the authors to either add a regularization term that prevents the α collapse or reposition Step 2 as an empirically tuned early-stopped heuristic supported by a sensitivity analysis. The discrepancy-functional inconsistency (α vs α²) is also central because it affects the training labels and the loss-balancing derivation. I lean toward major revision rather than rejection because the Step-1 acceleration idea is sound and the numerical infrastructure is substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it learns Tikhonov regularization parameters for a noisy operator (the LSM setting) by first training on low-resolution Morozov maps and then refining on the test data itself. That combination is not in the cited literature, and the problem it targets—fast, high-quality LSM imaging without manual threshold tuning—is real. Step 1 is well anchored: the Morozov-supervised pretraining gives the network an independent target, and the synthetic experiments show the informed loss generalizes better than the basic one. The proposed loss-weighting scheme, derived from GradNorm/DynScl without a separate optimization loop, is a solid engineering contribution, and the comparison with SoftAdapt, LRA, and ReLoBRaLo is honest and useful.\n\nThe soft spot is in Step 2, and it is not minor. The stress-test note is correct: for a fixed right-hand side, the Tikhonov loss evaluated at the Tikhonov minimizer is sum_j α |(u_j^*, u_L)|^2 / (α + d_j^2), which is strictly increasing in α and tends to zero as α → 0. Unconstrained minimization of the imaging objective therefore drives α toward zero, and both the training and validation losses share this degeneracy. The reported contrast gains after Step 2 are thus a property of the particular early-stopping thresholds in Algorithm 2 (σ_r = 5, σ_a = 10^-4) and the chosen training duration, not a proven consequence of minimizing J_img. The paper even acknowledges that Step 2 needs external regulation, but it does not offer a theoretical justification for why the stopping rule should select a good α rather than an overfitting one. I would want to see either a modified objective with an interior optimum (e.g., penalizing small α implicitly through the network parameterization) or a convincing empirical demonstration that the selected stopping point is stable across thresholds and noise realizations. There is also a concrete internal inconsistency: Eq. (15) and Eq. (32) define the discrepancy functional with α − η^2 D^2, but the derivative in Eq. (19) does not match that function. And Eq. (24) appears to sum over the wrong index. These are fixable but should be corrected.\n\nOther concerns: no code or data released, no error bars, and the final evaluation uses the same test patterns that drove Step 2 training, so the 'improvement' is partly self-referential. The application is synthetic only, though the test setup (elastic plate, evolving crack cloud) is more realistic than most toy examples.\n\nWho is this for? Researchers working on sampling methods, especially LSM, who want to automate regularization without ground-truth labels. The paper deserves a serious referee: the core idea is worth engaging with, and the flaws are identifiable and correctable. I would send it to review but insist that the authors address the Step 2 objective degeneracy, release code/data, and report sensitivity to the stopping thresholds before acceptance.","headline":"A worthwhile two-step learning idea for noisy-operator regularization, but the Step 2 quality gains are carried by early stopping, not by the stated objective.","tokens_in":26902,"tokens_out":2801,"would_cite":false,"duration_ms":31976,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F22","65J20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-step neural network learns regularization maps for inverse problems with noisy operators and improves image contrast.","keywords":["linear sampling method","Tikhonov regularization","Morozov discrepancy principle","regularization networks","inverse scattering","ultrasonic imaging","noisy operator","loss balancing"],"falsifier":"In the 25-percent-noise cases, re-run Step 2 with the stopping thresholds $\\sigma_a$ and $\\sigma_r$ removed so training continues for many more epochs, and record the defect-to-background contrast at every epoch: if contrast keeps improving after the rule would have stopped, or never peaks near the rule's stop point, then the image enhancement is not produced by the loss-trajectory regulation that carries the claim.","tokens_in":25806,"feed_emoji":"🔍","tokens_out":8536,"duration_ms":89281,"temperature":0.7,"pith_summary":"The paper is trying to show that the expensive per-pattern choice of a regularization parameter in linear sampling method imaging can be learned, even when the forward operator itself is noisy. The proposed scheme trains a small MLP on a low-resolution subset of right-hand side patterns to imitate Morozov discrepancy maps made with a deliberately rough threshold, then refines the network by minimizing the Tikhonov imaging objective under an early-stopping rule. The authors claim this two-step procedure accelerates regularization while improving image contrast, with the largest gains appearing in 25-percent-noise reconstructions of a complex evolving crack field. A sympathetic reader would care because, if the claim holds, real-time ultrasonic imaging of damage could bypass manual regularization tuning and use only a tiny fraction of the right-hand side data.","feed_headline":"Discrepancy-trained networks beat manual tuning in noisy crack imaging","feed_subtitle":"A two-step network learns regularization maps from a 0.35 percent subset of patterns and improves contrast at high noise.","key_machinery":"The load-bearing object is the regularization network (R-Net), an MLP whose input is the right-hand side pattern projected onto the left singular vectors of the noisy scattering operator, $U^*u_L$, and whose output is a Tikhonov parameter $\\alpha_{NN}$ for that pattern. The spectral filter factors $D_{jj}/(\\alpha_{NN}+D_{jj}^2)$ then define the regularized solution $g_{NN}=VD^\\dagger U^*u_L$, which is what the linear sampling method indicator uses. The argument is carried by two training mechanisms: (1) in Step 1, the informed loss combines a normalized misfit to the Morozov $\\alpha_M$ with the squared Morozov discrepancy functional $J_2^t$ evaluated at a rough threshold $\\eta_\\circ$, and the chosen weights in (19) make both gradient terms $O(\\partial \\alpha_{NN}/\\partial w)$, so the discrepancy logic is learned without an inner optimization for loss balancing; (2) in Step 2, minimizing $J_{img}$ tailors $\\alpha$ toward a min-norm solution, and Algorithm 2 stops training when the rms of normalized training or validation loss variations crosses a threshold or their ratio leaves an interval, which is the mechanism claimed to convert extra training into higher image contrast without overfitting.","core_discovery":"The central discovery claimed is that a network trained first to imitate cheap, low-resolution Morozov discrepancy maps and then fine-tuned on the Tikhonov imaging loss can produce linear sampling method reconstructions whose contrast matches or beats the Morozov maps obtained by per-time-step manual tuning, even when the scattering operator carries 25 percent noise. The network takes the projected right-hand side $U^*u_L$ as input and outputs the Tikhonov parameter $\\alpha$; with that $\\alpha$ the spectral filter factors $D_{jj}/(\\alpha + D_{jj}^2)$ supply the regularized solution. Step 1 uses an informed loss that includes the discrepancy principle itself, with adaptive weights that normalize the gradients without a second optimization loop, and Step 2 minimizes the Bayes-risk Tikhonov loss with an early-stopping rule based on normalized training-versus-validation loss trajectories. The evidence is a synthetic elastic-plate experiment in which a single crack grows into a cloud of fifteen cracks over five sensing steps; at 25 percent noise the Step-2 images show higher defect-to-background contrast than the manually tuned Morozov baseline at almost every sensing step.","pith_inferences":["Extension: Apply the same scheme to generalized linear sampling method or time-domain imaging, where per-pattern root-finding is even more expensive; the discrepancy functional changes but the projection-plus-early-stopping logic should transfer, and this is untested in the paper.","Extension: The only guard against $\\alpha$ collapsing toward zero in Step 2 is the stopping rule, so a deployment should monitor a ground-truth-free contrast metric alongside the loss trajectories; the paper reports no correlation analysis between loss rms and contrast.","Extension: If low-resolution-to-dense generalization is the mechanism, substituting a resolution-independent network for the fixed-size MLP could let one trained model serve arbitrarily fine grids, which the paper flags as possible but does not demonstrate."],"forward_implications":["Dense regularization maps for the full 100 by 100 search grid can be produced from a network trained on only 2,500 of 720,000 right-hand side patterns.","A discrepancy-informed Step 1 generalizes better than direct regression on Morozov maps, so the choice of training loss determines whether the network can be refined in Step 2.","Step 2 improves the mean and maximum contrast of LSM images relative to manually tuned Morozov reconstructions, with the largest gains appearing at 25 percent noise and late sensing steps.","Because Step 2 needs no labeled optimal regularization maps, the network can adapt to test data online, removing the need to tune the Morozov threshold by hand."],"supporting_citations":[{"why":"Introduces supervised bilevel learning of regularization functionals, the paradigm that Step 1 and Step 2 extend to noisy operators.","marker":"[10]"},{"why":"Provides the classical linear sampling method formulation whose Tikhonov loss is the imaging objective being regularized.","marker":"[38]"},{"why":"States the Morozov discrepancy principle that generates the low-resolution regularization maps used as Step 1 training targets.","marker":"[43]"},{"why":"Supplies the regularization theory and discrepancy-principle analysis that justify the loss terms in the informed training mode.","marker":"[45]"},{"why":"Gives the GradNorm gradient-balancing logic whose derivative-normalization goal the proposed adaptive weights replicate without inner optimization.","marker":"[46]"},{"why":"Contributes the dynamic-scaling loss-balancing idea that the paper adapts for its many-objective Step 1 loss.","marker":"[47]"},{"why":"Establishes the universal approximation property used to justify the MLP architecture of the regularization network.","marker":"[49]"},{"why":"Defines contrast metrics inspired by Weber contrast that the paper uses to compare Morozov and R-Net reconstructions.","marker":"[58]"}],"fun_headline_variants":["Neural network learns regularization for noisy inverse problems","Two-step training sharpens crack images under 25 percent noise","Discrepancy-informed training accelerates and enhances crack imaging","Two-step training: learn from discrepancy, refine with Tikhonov"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's claimed contrast gain in Step 2 rests on a heuristic early-stopping rule: the relative trajectories of training and validation Tikhonov losses are assumed to reveal when further minimization stops helping the image, and if that rule fails on a new geometry or noise level, the refinement step can make reconstructions worse rather than better.","fun_headline_variants_meta":{"raw":{"variants":["Neural network learns regularization for noisy inverse problems","Two-step training sharpens crack images under 25 percent noise","Discrepancy-informed training accelerates and enhances crack imaging","Two-step training: learn from discrepancy, refine with Tikhonov"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0011,"raw_usage":{"total_tokens":4651,"prompt_tokens":1068,"completion_tokens":3583,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":3516}},"tokens_in":684,"tokens_out":3583,"duration_ms":28155,"temperature":1.0,"reasoning_tokens":3516,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:44:53.210380+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the 25-percent-noise cases, re-run Step 2 with the stopping thresholds $\\sigma_a$ and $\\sigma_r$ removed so training continues for many more epochs, and record the defect-to-background contrast at every epoch: if contrast keeps improving after the rule would have stopped, or never peaks near the rule's stop point, then the image enhancement is not produced by the loss-trajectory regulation that carries the claim.","supporting_citations":[{"cited_title":"Cakoni, D","cited_arxiv_id":null,"evidence_quote":"Provides the classical linear sampling method formulation whose Tikhonov loss is the imaging objective being regularized."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"States the Morozov discrepancy principle that generates the low-resolution regularization maps used as Step 1 training targets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the GradNorm gradient-balancing logic whose derivative-normalization goal the proposed adaptive weights replicate without inner optimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the dynamic-scaling loss-balancing idea that the paper adapts for its many-objective Step 1 loss."},{"cited_title":"Avatavului, M","cited_arxiv_id":null,"evidence_quote":"Defines contrast metrics inspired by Weber contrast that the paper uses to compare Morozov and R-Net reconstructions."}],"review_version":1}