{"id":"7f26ab9a-3da9-4219-b6ff-bd819ba3e80e","arxiv_id":"2501.04608","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An empirical study proposing DeMUN, a memory-based unrolled network, and finding that intermediate loss and residual connections improve reconstruction while projector depth beyond five layers matters little.","lead":"This paper introduces DeMUN, an unrolled network that learns to combine gradient information from all previous iterations, and reports a large ablation study across losses, residual connections, network depth, noise, and measurement matrices. The main practical finding is that a simple default configuration, unweighted intermediate loss plus residual connections, works well and reduces the number of design choices practitioners must make.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run PSNR tables without seed variance leave the 'uniform' ordering claims unquantified; the central recommendation needs repeated-training confidence intervals before it can be called robust.","rationale":"The reader's conditional verdict is reasonable. The strongest claim is an empirical rank ordering, so the load-bearing condition is that the observed PSNR gaps are reproducible rather than artifacts of a single training run. The paper contains no error bars and no seed variation; this is the weakest link because the central 'uniformly outperform' statement covers every configuration, and several gaps are within the range of typical optimization stochasticity. The reader's stated weakest assumption, that results were obtained only on Gaussian and DCT matrices and may not transfer to structured real-world operators such as MRI or CT, is a legitimate scope limitation and is acknowledged in part by the paper. However, it concerns external validity rather than the internal comparison. This stress test therefore places more weight on the missing variance estimates. The recommendation is unchanged from the reader's conditional verdict: the paper should add repeated runs with variance reporting, soften overclaims based on point estimates, and ideally validate on a real forward model to strengthen the practical guidance.","tokens_in":24743,"tokens_out":7015,"duration_ms":67863,"concrete_test":"Retrain the configurations in Tables 5 and 6 (DeMUN and PGD, with and without residual connections, T=15 and T=30, m/n=0.1 and 0.4) and the ℓi,ω sweep in Table 8 using 5 independent seeds each, reporting mean ± std PSNR on the identical 2500-image test set. If every pairwise ordering (ℓi,1 over ℓll; residual over no-residual; DeMUN over PGD) holds with per-config std smaller than the minimum observed gap, the claim is supported; if any ordering reverses or std exceeds the gap, the 'uniformly' language must be softened. As a secondary reproducibility check, run the released GitHub implementation for one selected cell, e.g., Table 5, DeMUN residual, T=15, m=0.4n, and confirm that the PSNR matches the reported 34.43 within a stated tolerance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on pairwise PSNR comparisons that are directionally consistent but are reported from single training runs. Several gaps are very small: Table 5 gives PGD no-residual 26.42 vs residual 26.43 at T=15, m=0.1n; Table 7 shows ℓi,1 vs ℓi,0.95 differences up to roughly 0.2 dB; Tables 10-12 even show L=3 occasionally at or above L=5 in the same configuration (e.g., Table 10, m=0.1n: 26.33 vs 26.29; Table 11, m=0.4n: 34.49 vs 34.43). Adam training with random initialization and stochastic batching has run-to-run variation that can exceed these gaps, and the paper reports no seeds, no repeated runs, and no error bars. Since Hypothesis 1 and the conclusion use the words 'uniformly' and 'consistently,' the absence of variance estimates makes the strongest claim underdetermined. This is a correctness risk rather than a style issue: if retraining with different seeds reverses even a few of the small-gap comparisons, the claimed uniformity fails and the practical recommendation needs qualification. The external-validity limitation (Gaussian and DCT matrices only, no real MRI/CT forward operator) is real but secondary; it limits scope without invalidating the internal ordering.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies design choices for unrolled networks solving linear inverse problems y = Ax + w. It introduces DeMUN, an unrolled architecture whose linear step is a learned combination of the current iterate and all past gradient terms A^T(y - Ax_j), implemented as trainable 1x1 convolutions; this generalizes PGD, Nesterov, and AMP in a single framework. The core empirical contribution is an extensive ablation across loss functions (last-layer, weighted intermediate, and skip-L intermediate losses), residual connections, the number of unrolled steps T ∈ {5, 15, 30}, DnCNN projector depth L ∈ {3, 5, 10, 15}, sampling rates {0.1, 0.2, 0.3, 0.4}n, Gaussian and DCT measurement matrices, additive Gaussian noise levels, and image resolutions from 32x32 to 80x80. Based on test PSNR on 2500 ImageNet validation crops, the authors propose four hypotheses and recommend DeMUN trained with the unweighted intermediate loss ℓ_{i,1} and residual connections.","tokens_in":25033,"tokens_out":6209,"duration_ms":54722,"significance":"If the empirical ordering is reproducible, the paper would give practitioners a useful shortcut: they could avoid sweeping over optimization algorithms, loss functions, and projector depths when designing unrolled networks. DeMUN is a clean and simple generalization, the ablation grid is broad, the DCT and noise robustness checks go beyond many ablation studies, and the code is publicly linked. The main limitations are statistical and interpretive: all comparisons come from single training runs, several gaps are tiny (0.01–0.2 dB), the word 'uniformly' is used without error bars, and Hypothesis 4 is internally contradicted by the paper's own DCT results. With repeated-training statistics and refined claims, this would be a valuable reference for the field.","major_comments":[{"comment":"The central claims that intermediate loss and residual connections 'uniformly' or 'consistently' improve performance are based on single training runs. Many reported differences are smaller than typical Adam run-to-run variation: for example, Table 5 (T=15, m=0.1n) shows PGD without residual at 26.42 dB and with residual at 26.43 dB, and Table 4 (m=0.1n) shows DeMUN at 26.09 vs 26.29 dB. No seeds, repeated runs, or confidence intervals are reported, so the pairwise ordering is not quantified. Because Hypotheses 1 and 2 use universal language, please report multi-seed means with standard deviations (or at least the number of seeds and observed range) and, where gaps are below roughly 0.1 dB, state that the difference is not statistically meaningful. This is load-bearing: reversal of even a few small-gap comparisons would weaken the 'uniformly' claims.","section":"Sec. 4.3–4.4, Tables 4–6"},{"comment":"Hypothesis 4 is not robust as stated. The hypothesis says that increasing L from 3 to 5 yields a 'noticeable improvement,' but Table 10 shows L=3 outperforming L=5 at m=0.1n (26.33 vs 26.29 dB) and Table 11 shows the same at m=0.4n (34.49 vs 34.43 dB). Moreover, in Section 5.3 the authors acknowledge that for DCT matrices 'in most cases... the performance gain from increasing L from 3 to 5 is also marginal,' which directly contradicts the unqualified claim in Hypothesis 4. Please either weaken the hypothesis (for example, 'L=3 can be slightly worse or comparable to L=5; L=5 is a safe default') or provide statistical evidence for a genuine improvement from L=3 to 5.","section":"Sec. 4.6 and 5.3, Tables 10–12 and 18–20"},{"comment":"The paper states that 'Among all algorithms that we have unrolled... DeMUN achieves the highest PSNR values when trained with the intermediate loss,' and the tables support that restricted claim. However, the broad wording of Hypothesis 1 and the abstract-level summary can be read as claiming DeMUN superiority regardless of loss. Under the last-layer loss, Tables 1–3 show PGD and Nesterov sometimes beating DeMUN (for example, Table 2, m=0.2n: PGD 27.71 vs DeMUN 27.23; Table 2, m=0.1n: PGD 24.72 vs DeMUN 24.42). The footnote acknowledges this, but the main text should make the restriction explicit in the hypothesis statement and in Section 6. Please rephrase to say that DeMUN superiority is demonstrated for the intermediate-loss configuration and not claimed in general.","section":"Sec. 4.3, Hypothesis 1 and Tables 1–3"}],"minor_comments":[{"comment":"In the first bullet of Section 5.1, 'Tabel 13' should be 'Table 13'.","section":"Sec. 5.1"},{"comment":"The heading 'Loss F unction' contains an erroneous space; it should read 'Loss Function'.","section":"Sec. 2.1"},{"comment":"The additive noise is written as 'ω ∼ N(0, σ²I)', but ω already denotes the loss-weight parameter in Eq. (4); using a different symbol such as w would avoid confusion.","section":"Sec. 5.2"},{"comment":"The question 'Do we need more or fewer projections as we increase the number of projections?' should likely read 'as we increase the image resolution'.","section":"Sec. 5.4"},{"comment":"The intermediate-PSNR figures would be easier to read with labeled axes (projection index vs PSNR) and a legend; currently some figures are visually similar and rely only on the caption.","section":"Figures 4–13"},{"comment":"In the phrase 'theDeep Memory Unrolled Network' a space is missing after 'the'.","section":"Sec. 2.3"}],"recommendation":"major_revision","confidential_remarks":"This is a solid and unusually broad empirical study, but the missing repeated-run statistics is a real obstacle for a journal that expects quantitative claims. The 'uniformly' language in Hypotheses 1 and 2 needs to be backed by seed variance or explicitly relaxed. The contradiction between Hypothesis 4 and the DCT results in Section 5.3 must be resolved. If those points are addressed, the paper could be acceptable; the MRI/CT external-validity caveat is a scope limitation rather than an internal error."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a genuinely useful empirical paper for anyone building unrolled networks for linear inverse problems: it proposes a clean generalization of PGD/Nesterov/AMP, runs a broad ablation, and ships code. Second, the headline claims use 'uniformly' and 'consistently' where the evidence is single-run PSNR tables, and some of the gaps are smaller than the run-to-run noise of Adam retraining. That combination makes the paper a good candidate for peer review, but not for acceptance as-is.\n\nDeMUN's update in Eq. 6, with learnable coefficients over all past gradients, is a nice unification; PGD, Nesterov, and AMP are recoverable as special cases, AMP modulo the Onsager term, which is fine for an unrolled approximation. The ablation spans loss functions, residual connections, number of unrolled steps, projector depth, measurement matrices, noise, and resolution, and the main direction — intermediate loss plus residual connections helps — holds across most of the tables. The code link in Appendix A is a real plus. The authors' own limitation statements in Sections 4.6 and 5.3 are accurate as far as they go.\n\nThe biggest problem is the absence of any variance information. Every table is a single training run on a fixed test set. Several comparisons that ground the 'uniform' claims are tiny: Table 5 gives PGD 26.43 with residual versus 26.42 without at T = 15, m = 0.1n; Tables 10 and 11 show L=3 at or above L=5 in several cells (e.g., Table 10, m = 0.1n: 26.33 vs 26.29; Table 11, m = 0.4n: 34.49 vs 34.43). With Adam and random initialization, run-to-run PSNR differences can easily exceed 0.2 dB, so directionally consistent but small gaps do not support 'uniformly.' The fix is standard: repeat with several seeds, report means and error bars or confidence intervals, and soften the language accordingly.\n\nHypothesis 4 is the weakest part. The claim that increasing the projector depth from L=3 to L=5 gives a 'noticeable improvement' is not robust across the reported settings, and the authors essentially concede this for DCT matrices in Section 5.3 ('the performance gain ... is also marginal'). That hypothesis should be rewritten or dropped.\n\nLesser concern: external validity is limited to Gaussian and DCT matrices on grayscale center crops, with no real MRI or CT forward operator. The paper acknowledges this, so I read it as a scope limitation, not a fatal flaw. Circularity is not an issue: the hypotheses are empirical summaries checked on additional settings.\n\nWho is this for? Practitioners who want a default configuration for unrolled networks, and researchers working on algorithm unrolling. It deserves a serious referee: the architecture is real, the sweep is broad, and code is shipped. But it needs a revision with repeated runs and confidence intervals before the robustness claims can be trusted. Send it to peer review, with the expectation of major revision.","headline":"Useful broad ablation and a clean generalization (DeMUN), but no seed variance undercuts the 'uniformly' claims; accept for peer review, then revise.","tokens_in":25533,"tokens_out":6453,"would_cite":true,"duration_ms":52833,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A memory-based unrolled network that reuses every past gradient and is trained with an unweighted sum of intermediate reconstruction errors achieves the best reconstruction quality among the tested unrolled designs.","keywords":["unrolled networks","linear inverse problems","image reconstruction","deep memory unrolled network","intermediate loss","residual connections","ablation study","compressed sensing"],"falsifier":"Take a structured measurement operator from a real imaging modality, for example radial or Cartesian-sampled MRI with realistic receiver noise, and train the recommended configuration (DeMUN, unweighted intermediate loss, residual connections) against a standard unrolled baseline such as projected gradient descent with last-layer loss. If the recommended configuration fails to match or beat that baseline in reconstruction quality on held-out images, the paper's central recommendation would be shown not to transfer.","tokens_in":1725,"feed_emoji":"🖼️","tokens_out":1672,"duration_ms":89033,"temperature":0.7,"pith_summary":"This paper tries to turn the design of unrolled networks for linear inverse problems from a costly search into a fixed recipe. It introduces DeMUN, an unrolled network whose update at every step is a learnable combination of the current estimate and the gradient formed at every previous step, so it contains standard projected gradient descent, momentum acceleration, and approximate message passing as special cases. Through extensive ablations on Gaussian and DCT measurement operators with natural-image data, the paper argues that training DeMUN with the unweighted intermediate loss, the sum of squared errors over all unrolled estimates, together with residual connections yields the best Peak Signal-to-Noise Ratio among all designs tested. It also reports practical rules of thumb: loss weighting that keeps at least 0.75 weight on early steps works as well as equal weighting, and projector depth between 5 and 15 layers hardly matters. If the recipe transfers to real imaging operators, practitioners can skip most design decisions.","feed_headline":"Keep every gradient and supervise every step for best reconstructions","feed_subtitle":"One design recipe beats the usual unrolled-network choices, so practitioners can skip the search.","key_machinery":"The machine that carries the argument is the Deep Memory Unrolled Network (DeMUN) update $$\\tilde{x}_i = \\alpha_i x_i + \\sum_{j=0}^{i} \\beta_i^j A^\\top(y - A x_j),$$ followed by a learned projection $x_{i+1} = P_C(\\tilde{x}_i)$ or the residual variant $x_{i+1} = \\tilde{x}_i + P_C(\\tilde{x}_i)$. The coefficients $\\alpha_i, \\beta_i^j$ are learned, so the data chooses how much of each past gradient to keep; the update reduces to standard gradient descent when only the current gradient survives and to momentum- or message-passing-style updates under other coefficient patterns. The second piece is the unweighted intermediate loss $\\ell_{i,1} = \\sum_{i=1}^{T} \\|x_i - x^*\\|_2^2$, which supervises every unrolled estimate instead of only the final one. Together they make the training landscape smoother and let extra unrolled steps keep improving the estimate, which underlies the headline performance claims.","core_discovery":"The central discovery is that the choice of algorithm to unroll can be eliminated: DeMUN lets training learn the memory coefficients, and with an unweighted intermediate loss and residual connections it consistently outperforms unrolled projected gradient descent, momentum acceleration, and approximate message passing. With the last-layer loss, extra unrolled steps give almost no gain; with the intermediate loss, performance improves with more steps and plateaus rather than degrading. The paper also finds that weighting early layers slightly less (omega=0.95, 0.85) does not hurt, but weighting them much less (omega <= 0.5) or using skip-layer supervision degrades performance, and that increasing the number of projector layers from 5 to 15 is inconsequential while dropping to 3 hurts. These results hold for both Gaussian and DCT matrices, under additive noise, and across image resolutions from 32 by 32 to 80 by 80.","pith_inferences":["The same recipe could be stress-tested on structured medical operators such as radial or Cartesian-sampled MRI and fan-beam CT; if it holds there, the paper's recipe becomes a default starting point for new imaging systems.","The slower-than-input-SNR degradation suggests the trained network behaves partly as a denoiser; a testable extension is to compare DeMUN's reconstruction error against state-evolution predictions of approximate message passing in the large-system limit.","The interaction between DeMUN's many parameters and the last-layer loss, where DeMUN sometimes loses to projected gradient descent, hints that intermediate supervision rescues high-capacity unrolled networks from poor local minima; one could test this by visualizing loss-landscape curvature or varying parameter count."],"forward_implications":["Training with the unweighted intermediate loss turns extra unrolled steps from a liability into a monotone, plateauing improvement, so a practitioner can pick a large step count without fear of overfitting.","The algorithm-choice decision disappears: because DeMUN contains projected gradient descent, momentum, and message-passing-style updates as special cases, the learned coefficients select the effective update, and the paper finds DeMUN matches or beats those baselines.","The loss-weighting choice is forgiving: any omega in [0.75, 1] performs about as well as the unweighted intermediate loss, so users do not need to tune the exponential weighting precisely.","Projector capacity is not a sensitive dial: 5 to 15 convolutional layers give nearly identical PSNR, including under noise and DCT operators, so small projectors can be used to save computation.","Robustness checks show the same design choices work for DCT matrices, additive Gaussian noise, and resolutions from 32 to 80, and reconstruction degrades more slowly than the input SNR."],"supporting_citations":[{"why":"Introduced unrolled networks as iterations of optimization algorithms with learned projections, the class of architectures this paper studies.","marker":"[40, 41]"},{"why":"Supplies the convolutional denoiser architecture used as the learned projector in every unrolled network tested.","marker":"[51]"},{"why":"Motivated the intermediate-loss family by showing that supervising intermediate outputs improves reconstruction tasks.","marker":"[48]"},{"why":"Provided the approximate message passing algorithm and its learned variant used as a comparison baseline for DeMUN.","marker":"[6, 15]"},{"why":"Gave the accelerated first-order method that is unrolled as the momentum baseline in the comparison.","marker":"[45]"},{"why":"Established residual connections as a training stabilizer, the mechanism Hypothesis 2 adds to DeMUN.","marker":"[52, 53]"}],"fun_headline_variants":["Learn memory coefficients, use intermediate loss for best unrolled nets","Skip the design hunt: unrolled net recipe discovered","One recipe beats unrolled PG, momentum, and AMP","Unrolled nets: learn memory, keep losses, get best results","Best unrolled networks: learn memory, weight early layers lightly"],"cache_read_input_tokens":27648,"weakest_assumption_plain":"The recipe was measured on Gaussian and DCT measurement operators acting on grayscale natural-image crops up to 80 by 80, and the paper's recommendations assume these findings transfer to other linear inverse problems, such as MRI or CT, where the forward operator is structured, the noise enters differently, and the image statistics differ.","fun_headline_variants_meta":{"raw":{"variants":["Learn memory coefficients, use intermediate loss for best unrolled nets","Skip the design hunt: unrolled net recipe discovered","One recipe beats unrolled PG, momentum, and AMP","Unrolled nets: learn memory, keep losses, get best results","Best unrolled networks: learn memory, weight early layers lightly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1395,"prompt_tokens":932,"completion_tokens":463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":379}},"tokens_in":548,"tokens_out":463,"duration_ms":4776,"temperature":1.0,"reasoning_tokens":379,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:27:59.679647+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a structured measurement operator from a real imaging modality, for example radial or Cartesian-sampled MRI with realistic receiver noise, and train the recommended configuration (DeMUN, unweighted intermediate loss, residual connections) against a standard unrolled baseline such as projected gradient descent with last-layer loss. If the recommended configuration fails to match or beat that baseline in reconstruction quality on held-out images, the paper's central recommendation would be shown not to transfer.","supporting_citations":[{"cited_title":"Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising","cited_arxiv_id":null,"evidence_quote":"Supplies the convolutional denoiser architecture used as the learned projector in every unrolled network tested."}],"review_version":1}