{"id":"402d94cb-db5b-48d6-b120-af856ffa5b79","arxiv_id":"2509.15494","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A wavelet-based multi-scale neural network with a local kernel enhancement module achieves better rate-distortion on scientific datasets than standard implicit neural representations.","lead":"Scientists propose WIEN-INR, a neural-network method for compressing large experimental datasets by splitting them into different frequency scales with wavelets and adding a module that recovers fine details. Early tests on X-ray and neutron scattering data show higher fidelity at the same model size than existing neural compressors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Enhancement module's zerotree prior is load-bearing; gains may not transfer to data without parent–child wavelet correlation, so 'data-agnostic' overclaims.","rationale":"The reader's weakest_assumption identifies the same mechanism I see as load-bearing: the enhancement module in Eq. 5 assumes fine-scale wavelet coefficients are locally predictable from coarser-scale coefficients. I agree with that identification. The paper is otherwise coherent: the multi-resolution decomposition, independent SIRENs per scale, and the enhancement module are clearly described, and the Wavelet-INR vs WIEN-INR ablation is informative evidence for the module's contribution on the tested data. The three scientific datasets are relevant and the rate–distortion comparisons are plausible. However, because the enhancement module is the differentiator and it embodies a strong prior, the absence of any test on data without that prior leaves the 'data-agnostic' and 'full spectrum' conclusions unsupported. This does not invalidate the scoped empirical claim about the three datasets, so the reader's CONDITIONAL verdict remains appropriate rather than moving to ACCEPT or REJECT. The main adjustment is to require, as a condition of acceptance, either a demonstration on data lacking cross-scale correlation or a revision of the generalization claims.","tokens_in":11696,"tokens_out":6181,"duration_ms":67530,"concrete_test":"Construct a synthetic 3D volume y = s + n, where s is a smooth field (e.g., Gaussian-blurred random blobs) and n is spatially independent Gaussian noise with a controlled SNR. Train WIEN-INR and a same-parameter-count SIREN (and, if possible, Wavelet-INR) at two or three compression ratios using the paper's protocol (SIREN scales, J=4, Haar, nr=3, half-precision storage). Compare PSNR at matched compression ratios. If WIEN-INR does not exceed SIREN when noise dominates, the enhancement prior is the source of the reported gains and the 'data-agnostic' claim must be scoped to datasets with cross-scale correlation. A real-data variant is to repeat on raw detector noise frames from the same instruments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central fidelity claim rests on the finest-scale enhancement module (Sec. III.c, Eq. 5). For each fine coordinate x, WIEN-INR predicts the detail coefficient d^i_{j-1}(x) as an inner product of (i) an r×r patch of the upsampled coarser-scale prediction φ^i_j and (ii) kernel weights output by a small SIREN P^i_{j→j-1,ξ}(x). This is a spatially varying linear predictor, so the fine band can only be recovered insofar as it is locally predictable from the coarser band. That is exactly the zerotree/geometric-regularity prior: significant fine coefficients occur where coarse coefficients are significant, and values are locally correlated across scales. For signals whose finest-scale coefficients are independent of the parent scale (i.i.d. noise, detector shot noise, textures with no cross-scale redundancy), the module cannot produce the missing high-frequency innovation; its optimum is a smoothed estimate and the claimed PSNR advantage should collapse. The paper evaluates only three speckle-dominated scattering datasets, which plausibly share this cross-scale structure, so the empirical claim is not yet shown to generalize. The paper's conclusion explicitly claims 'data-agnostic applicability' and 'full spectrum' representation; that is the part put in doubt. This is an external-validity concern, not an internal inconsistency: within the tested datasets the reported results can stand.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WIEN-INR, a hierarchical implicit neural representation for scientific data compression. A discrete wavelet transform partitions the data into approximation and detail sub-bands; each band is modeled by a compact SIREN, with the finest detail band predicted by an enhancement module that applies a coordinate-dependent kernel to the upsampled output of the next-coarser scale. Experiments on three scattering datasets (X-ray diffraction, neutron scattering, ultrafast X-ray scattering) report higher PSNR at matched compression ratios than FE, SIREN, WIRE, and TCNN, together with robustness to wavelet basis and decomposition depth.","tokens_in":12080,"tokens_out":11285,"duration_ms":106184,"significance":"If the reported results are reproducible, WIEN-INR offers a practical improvement to INR-based compression for scientific data. The modular design (wavelet decomposition as preprocessing, optional enhancement module) is clean and the ablation against Wavelet-INR in Fig. 6 is valuable evidence that the enhancement module adds representational capacity beyond the multi-scale decomposition itself. The paper also includes a useful robustness analysis (Fig. 7) for wavelet type, number of scales, and window size. No code is provided and no theoretical guarantees are claimed, so the contribution is empirical.","major_comments":[{"comment":"The sums run over j=1..J−1. With the indexing used in the paper (j=1 finest, j=J coarsest; see Sec. III.c where the enhancement module predicts d1 from d2), a J-level DWT has J detail sub-bands d1..dJ. As written, the coarsest detail band d_J is omitted from the training objective. This is load-bearing: it changes which coefficients the networks are trained to represent. Please correct to j=1..J, or precisely define a different indexing convention and apply it consistently.","section":"Sec. III.a, Eqs. (2) and (3)"},{"comment":"The central rate-distortion claim rests on single-run curves. No number of seeds, error bars, or statistical significance tests are reported. Since the differences between methods at a given compression ratio appear modest and the baseline precision is not specified, please provide multiple independent runs and report mean ± std, or at least state the seed variability.","section":"Sec. IV.b, Fig. 4"},{"comment":"The conclusion claims 'data-agnostic applicability,' but the enhancement module (Sec. III.c, Eq. 5) assumes fine-scale wavelet coefficients are locally predictable from coarser-scale coefficients (a zerotree-like prior). The paper itself states in Sec. IV.a that applying the enhancement to the finest band is a data-dependent choice. All three evaluation datasets are scattering measurements with speckle/texture, which plausibly share this cross-scale structure. To support the 'data-agnostic' claim, either test on data without such parent-child correlation (e.g., i.i.d. noise, natural images) or temper the conclusion to specify the applicable data regime.","section":"Sec. VI vs Sec. IV.a"}],"minor_comments":[{"comment":"The notation 'N∇(x)' should be 'N_r(x)' for the r×r neighborhood. Also clarify whether the patch is extracted from φ^i_j or its upsampled version φ^i_{j−1}; the current text and equation are inconsistent.","section":"Sec. III.c, Eq. (5)"},{"comment":"The caption says 'coarse levels (a and d1)', but elsewhere d1 is the finest detail band (e.g., Sec. III.c describes enhancing 'd2 to d1'). Please clarify which scale is meant, or correct the indexing.","section":"Fig. 4a caption"},{"comment":"The enhancement window size is given as '(2^1+1)^p to (2^3+1)^p'; this is a roundabout way to say 3^p to 9^p. Please simplify for readability.","section":"Table I"},{"comment":"The description 'post-training quantization of network weights (division by two)' is unclear. Division by two is a scaling operation, not quantization to half precision. Please specify the actual quantization/rounding scheme used.","section":"Sec. V"},{"comment":"Reference [30] appears to duplicate reference [7] (both cite the same paper by Lu et al. on compressive neural representations of volumetric scalar fields). Please consolidate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the journal. The main technical concern is that the reported gain of the enhancement module over Wavelet-INR may be dataset-dependent, and the 'data-agnostic' claim in the conclusion goes beyond the evidence. I would encourage the editor to request corrected Eqs. (2)/(3) and additional seed runs before acceptance. The missing appendices (A1, B1, C1) should also be checked during review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat's actually new here: WIEN-INR combines a multi-level wavelet pyramid (per-scale SIRENs) with a lightweight enhancement module that predicts the finest detail band by applying a learned local kernel to the upsampled coarser-scale prediction. The ablation against Wavelet-INR (their own multi-scale baseline) gives real evidence that the enhancement module contributes, and they honestly disclose the concurrent CF-INR work. The method is coherent, physically motivated for speckle data, and robust across wavelet bases and depth settings. That is worth having.\n\nThe soft spots are mostly about verification, not about the core idea. No code or data; curves are single-run with no error bars; baseline storage precision is unstated. Those are fixable in a revision. The more substantive concern is the enhancement module's prior: fine-detail coefficients are predicted as a local linear function of the coarse-scale prediction. That works only when there is parent–child correlation across scales, the zerotree-type regularity that speckle-dominated scattering data plausibly has. For datasets whose finest-scale content is spatially independent or uncorrelated with the coarser band (shot noise, many textures), the module cannot synthesize the missing innovation, and the claimed PSNR advantage should shrink or vanish. The paper itself concedes the enhancement placement is data-dependent, yet the conclusion claims \"data-agnostic applicability.\" That mismatch needs to be either removed or properly scoped. Within the three tested datasets the results can stand; the overclaim is in the generalization rhetoric, not in the reported numbers.\n\nThe math is coherent modulo typos (e.g., Eq. 5), and the citation pattern to wavelet coding and zerotree coding is appropriate—this is not a case of missing prior art. The paper is a solid empirical contribution to INR-based scientific data compression, likely to be useful to anyone working on X-ray/neutron scattering data workflows.\n\nI would send this to peer review. The core architecture is worth refereeing even with the external-validity caveat, and the reproducibility gaps are addressable. My own verdict would be conditional: require code/data or at least multi-seed results, and tone down the data-agnostic claim. A serious referee should be able to push it to accept.\n\nBest,\n[You]","headline":"A genuinely useful multi-scale INR architecture for scientific data, with a real external-validity caveat: the enhancement module assumes cross-scale wavelet correlation, so 'data-agnostic' overclaims.","tokens_in":12528,"tokens_out":1317,"would_cite":false,"duration_ms":16160,"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":"By splitting signals into wavelet scales and adding a kernel-prediction network at the finest band, WIEN-INR lets compact neural representations keep fine speckle detail that standard INRs blur, with higher PSNR at matched compression.","keywords":["implicit neural representation","wavelet transform","multi-resolution analysis","scientific data compression","rate-distortion","X-ray scattering","neutron scattering","enhancement network"],"falsifier":"Take a real dataset, apply the wavelet transform, and randomly permute the spatial positions of the finest detail coefficients while keeping their marginal statistics intact; train WIEN-INR and the same framework without enhancement. If the enhancement no longer improves PSNR (or worsens it), the local coarse-to-fine correlation assumption is exposed as the source of the gain. A cleaner version: generate a volume whose finest-scale wavelet coefficients are independent random noise—there, the module has nothing predictable to learn and should show no rate-distortion benefit.","tokens_in":11639,"feed_emoji":"🔬","tokens_out":7720,"duration_ms":81488,"temperature":0.7,"pith_summary":"WIEN-INR is an attempt to get compact implicit neural representations to reproduce the full frequency content of scientific measurements, including sharp speckles and textures that small networks usually over-smooth. It applies a multi-level wavelet transform to split the data into frequency bands, trains a small coordinate network per band, and adds a lightweight enhancement module that predicts the finest detail band from an upsampled coarser prediction using a learned local kernel. On three experimental datasets—wide-angle X-ray diffraction, inelastic neutron scattering, and ultra-fast X-ray scattering—the authors report higher PSNR and SSIM at matched compression ratios than four standard INR baselines, along with preservation of the wavelet coefficient distribution. If correct, the method offers a practical way to store and query large scientific volumes as network weights without losing the fine structure that carries physical information.","feed_headline":"Split data by scale, then sharpen the finest band to beat bigger INRs","feed_subtitle":"WIEN-INR trains one network per wavelet band and a learned kernel that restores the fine detail larger models blur.","key_machinery":"Enhancement module (learned local deconvolution). For each orientation and scale transition, a lightweight SIREN takes a coordinate x and emits an r^p kernel; the kernel is applied as an inner product to an r×r patch extracted from the upsampled prediction of the next-coarser INR, and the result is trained to match the true wavelet coefficient at x. This is the mechanism that carries the claim of recovering subtle detail without inflating model size; it exploits the geometric regularity of wavelet coefficients—large amplitudes cluster along contours and textures, so fine detail is locally predictable from coarse structure. The multi-resolution preprocess (DWT plus one shared INR per scale fo","core_discovery":"WIEN-INR claims that the low-frequency bias of compact INRs can be overcome not by enlarging the network, but by reorganizing the representation: a discrete wavelet transform separates the signal into a coarse approximation and orientation-specific detail bands; each band is assigned its own small SIREN, with frequency parameters matched to the band; and the finest band is synthesized by an enhancement network rather than learned directly. The enhancement up-samples the next-coarser scale's prediction and applies a per-location r×r kernel—outputs of a separate lightweight network—to the local patch, matching the true finest coefficients. Across three raw experimental datasets, this yields hi","pith_inferences":["If the local coarse-to-fine correlation the enhancement assumes holds broadly, WIEN-INR effectively learns a data-driven zerotree-style prior for INR compression; combining it with entropy-coded weight quantization could push scientific INR codecs close to wavelet transform-codec rate-distortion limits.","The same coarse-to-fine predictor suggests a super-resolution mode the paper does not test: train only coarse scales, then let the enhancement module synthesize finer detail at inference, yielding resolution-scalable scientific data.","On data whose finest band is dominated by detector noise rather than geometric regularity, the enhancement could fabricate texture rather than recover signal; a synthetic white-noise-detail experiment would delineate the regime in which WIEN-INR should be used.","The reported robustness across wavelet bases masks a tunable resource-allocation question: how to split the parameter budget across scales. An adaptive rule based on per-band energy could extend the method to datasets with very different spectra."],"forward_implications":["Raw experimental volumes can be replaced by a set of small network weights, with decoding as a forward pass, and at a fixed weight budget the reconstructed data retains more fine-scale structure than with a single large INR.","Because the objective separates by scale, subnetworks can be trained in parallel; only the enhancement step depends on the previous scale, so encoding time can drop relative to monolithic INR training.","Each scale is a continuous coordinate function, so users can decode arbitrary regions of interest or resolutions without retraining, and the finest band is enhanced rather than emulated from scratch.","The enhancement module is modular, so existing coordinate-based INR pipelines can adopt it without changing their core architecture.","Preserving the wavelet-coefficient distribution, not just pixel error, means compression preserves multi-scale statistics that downstream physical analyses rely on."],"fun_headline_variants":["Wavelet-split INRs sharpen fine detail with tiny nets","Small INRs match big ones by splitting scales and enhancing","Enhance, don't enlarge: wavelet bands make INRs sharper","Per-band SIRENs plus a sharpening net beat bigger INRs","Tiny INRs recover fine detail via wavelet band enhancement"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that fine-scale wavelet detail is locally predictable from the next-coarser scale through a small learned kernel; if a dataset's finest structures are spatially independent of coarser structure, the enhancement module cannot recover them and the central fidelity gain disappears.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet-split INRs sharpen fine detail with tiny nets","Small INRs match big ones by splitting scales and enhancing","Enhance, don't enlarge: wavelet bands make INRs sharper","Per-band SIRENs plus a sharpening net beat bigger INRs","Tiny INRs recover fine detail via wavelet band enhancement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000134,"raw_usage":{"total_tokens":946,"prompt_tokens":684,"completion_tokens":262,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":428,"completion_tokens_details":{"reasoning_tokens":183}},"tokens_in":428,"tokens_out":262,"duration_ms":3744,"temperature":1.0,"reasoning_tokens":183,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:13:28.775781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real dataset, apply the wavelet transform, and randomly permute the spatial positions of the finest detail coefficients while keeping their marginal statistics intact; train WIEN-INR and the same framework without enhancement. If the enhancement no longer improves PSNR (or worsens it), the local coarse-to-fine correlation assumption is exposed as the source of the gain. A cleaner version: generate a volume whose finest-scale wavelet coefficients are independent random noise—there, the module has nothing predictable to learn and should show no rate-distortion benefit.","supporting_citations":[],"review_version":1}