W1. Table 2 cannot be reconciled with the authors' own supplementary results, and those results contradict the paper's central claims.
Table 2 is the paper's only evidence that the background representation, routing and prototype memory matter. The supplement contains the outputs of the authors' ablation scripts (paper_experiments/ablation_v9_seed0/routing/routing_ablation_table.csv and .../context/context_ablation_table.csv, with audit JSONs recording checkpoint hashes and configs). Every row of Table 2 equals a supplementary row after relabeling and/or adding round-number offsets (ΔSNR / CC / RMSE / ST-MAE):
| Table 2 row (paper) | Supplementary row | Paper − supplementary |
|---|
| PCD-Net: 7.8554 / 0.7667 / 0.0402 / 0.0269 | "w/o prototype memory": 7.8554 / 0.7667 / 0.0402 / 0.0269 | 0 / 0 / 0 / 0 |
| w/o Prototype: 7.1653 / 0.7032 / 0.0406 / 0.0272 | "Full PCD-Net": 7.7653 / 0.7632 / 0.0406 / 0.0272 | −0.6 / −0.06 / 0 / 0 |
| w/o Routing: 6.7970 / 0.6919 / 0.0605 / 0.0471 | "w/o routing (uniform alpha)": 7.7970 / 0.7619 / 0.0405 / 0.0271 | −1.0 / −0.07 / +0.02 / +0.02 |
| w/o Background: 5.7306 / 0.6628 / 0.0607 / 0.0474 | "No background" (all-zero context): 7.7306 / 0.7628 / 0.0407 / 0.0274 | −2.0 / −0.10 / +0.02 / +0.02 |
| w/o Decoder Adapter: 7.1170 / 0.7626 / 0.0406 / 0.0274 | "weighted-sum + concat fusion": 7.7170 / 0.7626 / 0.0406 / 0.0274 | −0.6 / 0 / 0 / 0 |
Each Table 2 row is compatible with exactly one supplementary model under this relation, and the match is unique even when only the ΔSNR column is used. Results from an independent run would agree in the last three ΔSNR digits with probability of roughly 10−11 for all five rows.
A label mix-up inside the pipeline seems excluded, for four reasons:
evaluate_routing_ablation_v9.py binds labels to the checkpoint arguments and validates each checkpoint's config against its label.
- The "Full PCD-Net" row is bit-identical in ΔSNR, CC and RMSE (e.g., 7.765267381063327) to the "Correct background" row of the context ablation, which evaluates only the full checkpoint.
- The same numbers reappear as the learned-τ point of Fig. 6's τ row.
- They coincide with the PCD-Net numbers of the main comparison (
global_compare.csv, 7.7652 dB), which match Table 1 and Fig. 4.
The inconsistency is visible even from the PDF alone. The "PCD-Net" row of Table 2 (7.8554 dB) lies 7.5 standard deviations above the five-seed PCD-Net result in Table 1 (7.7548±0.0134 dB). The actual full model in the supplement (7.7653 dB) agrees with Table 1.
Taken at face value, the authors' files show that zeroing or randomizing the background at inference changes ΔSNR by −0.035 dB, uniform routing gives +0.03 dB, removing the prototype memory gives +0.09 dB (the best variant), and weighted-sum concatenation instead of cross-attention gives −0.05 dB. All effects are below 0.1 dB, versus the 0.7–2.1 dB drops claimed in Table 2, and two of them have the opposite sign. No decoder adapter appears anywhere in the released code, so I could not see how a "w/o Decoder Adapter" variant was evaluated.
Fig. 6 is inconsistent even within itself. The selected configuration (K=16, τ=0.3) appears in both rows. In the top (K) row it reaches ≈9.1 dB ΔSNR and CC ≈0.81. In the bottom (τ) row the same configuration reaches 7.765 dB and CC 0.7632, which equals the supplementary full model and agrees with Table 1. In addition, K=8 is reported at ≈3.2 dB although removing the prototypes altogether yields 7.86 dB. According to routing_sensitivity_v9.py, the τ row is an inference-time sweep of a single checkpoint (it spans only ≈0.02 dB), and τ is in fact a learnable parameter (learnable_routing=True). The K results, by contrast, are read from an external CSV that is not included. The same script's overlay plot labels that series "K sweep (V7 Top-M)" (lines 121, 142), which suggests that the K row comes from an earlier model variant with hard Top-M selection rather than from the PCD-Net evaluated elsewhere in the paper.
Hence the statements in Sec. 4.2 (lines 411–421), and the abstract's claim that "the learned environment representations improve restoration", are not supported by the submitted evidence. This weakness alone is decisive for my rating.
W2. The cross-domain / parameter-efficient adaptation contribution is unsubstantiated and inconsistent with the released implementation.
- Tables 3–4 compare only PCD-Net variants. No baseline is adapted to the target domains, and no standard PEFT alternative (adapters, LoRA, BitFit, partial fine-tuning of a baseline) is compared, so it cannot be judged whether the "environment representation" makes the model more transferable than an ordinary network. Full fine-tuning is best on both target domains, and trainable-parameter counts are never reported, so "parameter-efficient" is not quantified.
- The paper describes a bottleneck decoder adapter (Eqs. 14, 26) with the encoder and decoder frozen. The released adaptation code (
transfer_v9.py) offers context_encoder, environment, decoder_only and environment_decoder; the last two unfreeze all decoder blocks (dec4–dec0, out_head), and no adapter module exists. It is unclear which configuration produced "Decoder Adjustment" and "Selective Adaptation".
- In the released adaptation scripts (defaults of
run_adaptation_ablation_v9.ps1 and the README example), target-domain events are corrupted with the STEAD noise bank (wave1.hdf5) for both adaptation and evaluation. The additive-noise distribution is then identical in source and target, so the shift is in the signal rather than in the "noise environment" that motivates the method.
- The adaptation entry points accept only
stead and mining. No file refers to the industrial dataset of Table 3, and no result files for Tables 3–4 are provided.
- Neither target dataset is described. Missing details include the source, the size and the sampling rate (the split script admits 50 Hz and 100 Hz mine data, while inputs are "6000 samples at 100 Hz"). The paper also does not say how clean references are obtained or how many target samples are used. The ethics statement says all datasets are public, but no source is given for either target dataset.
Secondary concerns (these would remain even if W1/W2 were resolved)
- Test vs. validation. All reported STEAD numbers are computed on
v7_seed0_val.csv. This file is exactly wave2.csv.sample(frac=0.1, random_state=0): the same 20,000 traces in the same order, i.e. the seed-0 validation split of the training pipeline. The PCD-Net checkpoint config has val_frac=0.1, seed=0 and val_samples=2000, and the best checkpoint is kept by validation CC. K and τ were "selected according to validation performance", and Fig. 6's τ row reproduces the reported test value (7.765 dB) at the learned τ. There appears to be no untouched test split.
- "Five random seeds." The supplement contains seed-0 runs only. Bandpass and Wavelet are deterministic yet show non-zero std in Table 1, which implies the test data change across seeds. In that case the reported stds should be close to the sampling standard error, ≈0.09–0.15 dB from the bootstrap CIs in
paper_model_comparison_table.csv. Instead they are 10–50× smaller (e.g., 0.0074 dB for Wavelet, 0.0030 dB for DPRNN).
- Baselines.
- DPRNN and DeepDenoiser are far worse than the identity map at high input SNR (ΔSNR −6.9 and −3.4 dB in the 5–10 dB bin; Fig. 4 and
per_group_compare.csv), which suggests under-training or misconfiguration.
- The code describes DeepDenoiser as a time-domain U-Net, whereas the original estimates STFT masks.
- The baseline implementations (
v3/...) are not included.
- PCD-Net has the most parameters (3.95M vs. 0.91–2.63M), a column that was dropped from Table 5.
- Simple strong baselines are missing: (a) the same backbone without conditioning, or with directly concatenated/FiLM background features (the "w/o prototype" variant, which feeds the background embedding as a single attention token, is in fact the best model); (b) a Wiener/spectral-subtraction filter using the pre-event noise PSD; (c) the background-conditioned GDiffuSE-/CDiffuSE-style baselines that are implemented in
compare_v9.py but not reported.
- The background is likely part of the input. STEAD traces are exactly 6000 samples with a median P arrival at sample ≈700. The 400-sample pre-P "background" therefore seems to be a sub-window of the 60 s input x itself, carrying no information unavailable to the baselines. This is consistent with the near-zero context ablation.
- Undisclosed objective / reproducibility.
- The training loss is never described; the configs combine MSE, frequency, gradient, "detect", quality-head and intermediate losses with event-region weighting.
- The reproducibility statement points to "the full training objective, the quality-aware training target, and the adaptive inference refinement procedure" in Appendix A, but none of these is there.
- The stated batch size (2) differs from the configs (4 for PCD-Net, 8 for the baselines).
- Figs. 5 and 7 are illegible.
- Related work. None of the following is discussed:
- Conditioning a denoiser on noise-only segments is well established in speech enhancement: noise-aware training (Seltzer et al., ICASSP 2013; Xu et al., Interspeech 2014), dynamic noise embeddings with adaptation (Lee et al., APSIPA ASC 2020), and noise-model guidance fitted on short noise-only segments (GDiffuSE, Yanir et al., 2025, arXiv:2510.04157).
- Degradation-embedding-weighted prompt/prototype components appear in all-in-one image restoration (AirNet, CVPR 2022; PromptIR, NeurIPS 2023).
- Eq. 26 is the standard bottleneck adapter (Houlsby et al., ICML 2019).