- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper proposes iPSD (Intelligent Partitioning for Self-supervised Denoising) and a zero-shot bandit variant iPSD-Zero for training deep single-channel EEG denoisers without clean reference signals. Rather than splitting a noisy signal into odd and even samples via fixed interleaved downsampling, iPSD segments into non-overlapping windows of length and uses a bidirectional GRU policy network trained via PPO to select one of 35 equal-size sub-signal splits per window, using the negative converged Noise2Noise loss of a 3-layer 1D ConvNet denoiser as the reward. iPSD-Zero restricts the partition to be identical across all windows of a single test signal and uses the lil'UCB bandit algorithm across the 35 candidate partitions. Experiments are reported on semi-synthetic CHB-MIT scalp EEG corrupted with WGN and surface EMG, and on a real-world wearable in-ear EEG sleep staging dataset.
Critical strengths
- Practical Application Motivation & Wearable Sleep-Staging Evaluation: Single-channel EEG artifact removal without clean ground-truth targets is a well-motivated problem for wearable neurotechnology, and evaluating denoising via downstream 5-class sleep staging on concurrent viscoelastic in-ear EEG and clinical scalp PSG is a relevant real-world testbed.
Critical weaknesses
- Unspecified Inference Procedure & Contradiction Between Figure 1, Eq. (2), and Eq. (6):
The paper never defines how the denoised signal is computed at test time, presenting three mutually incompatible descriptions:
- Figure 1 (Page 2) shows split into , mapped to , and interleaved into . Yet in Eq. (2), is trained to predict (at indices ) and predicts (at indices ), leaving unspecified whether is placed at or .
- Equation (6) (Page 5) defines the consistency loss using and , which applies directly to the full-length unpartitioned signal (where Mansour & Heckel, 2023 define ), contradicting Figure 1.
- Algorithm 1 & Appendix K (Page 24) imply that only the policy transfers across subjects and is trained from scratch ( epochs) on each test segment in (Appendix F, line 854), yet neither Algorithm 1 nor Algorithm 2 ever outputs .
- The
iPSD-ZeroBandit Paradox Refutes the Core Non-Stationarity Premise & Wastes of Compute:- Static Partitioning Beats Learned Window-Varying Policies: Sections 1 and 2.1 argue that EEG non-stationarity requires a sequential RL policy (
iPSD-GRU, params) to adapt partitions across all 320 local windows. YetiPSD-Zerouses a single static 8-sample partition repeated identically across all 320 windows—and outperforms every learned window-varying policy in Table 5 exceptiPSD-GRU(iPSD-MLP,iPSD-RNN,iPSD-U-Net, andDRin Table 2), matchesiPSD-GRUwithin on WGN, and produces wearable EEG outputs "visually indistinguishable from those of iPSD" (page 7, line 368). - Pointless
lil'UCBExploration Over Deterministic Arms: For a fixed test signal and arm , the split is 100% deterministic (the only variance across pulls is random weight initialization of , , line 840). Algorithm 2 (page 16) already pulls all 35 arms once at initialization (), and then runslil'UCBwith until pulls of the winning arm ( full 200-epoch trainings of ). Re-training over 300 times on the exact same deterministic signal pair wastes of compute compared to selecting the best arm after the 35 initial pulls.
- Static Partitioning Beats Learned Window-Varying Policies: Sections 1 and 2.1 argue that EEG non-stationarity requires a sequential RL policy (
- Selection / Collider Bias Invalidates Theorem 1 & Appendix C (Eqs. 7–11), and the "34–43% WGN Noise Correlation Reduction" Claim (Table 10) is a Statistical Fallacy:
- Breakdown of : Because observes the entire noisy signal before choosing , conditioning on the chosen partition induces selection/collider bias between and : (e.g., sorting pure noise within each window makes and strongly correlated, minimizing N2N MSE by predicting noise). Moreover, zero MMSE residual (Eq. 11) does not imply (in Appendix B's example , has zero MMSE), Theorem 1 subtracts from , and Appendix F (line 863) miscalculates as .
- WGN Correlation Fallacy (Table 10 & Page 9): Under i.i.d. WGN with samples, fixed interleaved downsampling (
ID) already produces 100% independent Gaussian vectors () whose finite-sample has exact theoretical expectation —matchingID(). Achieving on i.i.d. Gaussian noise is not "greater independence"; it proves that actively selects partitions conditioned on the noise realization to cancel sample inner products, violating noise independence.
- Mathematical Redundancy of PSNR with SNR, 7-Sample Receptive Field Paradox, and Parameter Discrepancy:
- : Subtracting Eq. (14) from Eq. (15) cancels identically: across every table. Claiming PSNR "emphasizes robustness to large transient spikes" (line 258) is mathematically false, and
Optimal-WT(EMG ) in Tables 1 & 7 has a transcription error (). - Receptive Field & Parameter Contradictions: (3-layer Conv1D, kernel size 3) has a 7-sample () receptive field—unable to span a single cycle of delta/theta/alpha rhythms (contradicting line 268)—and Section 3 claims parameters whereas Table 12 reports
0.33 m() parameters (a discrepancy).
- : Subtracting Eq. (14) from Eq. (15) cancels identically: across every table. Claiming PSNR "emphasizes robustness to large transient spikes" (line 258) is mathematically false, and
- Semi-Synthetic Ground-Truth Contamination, Log-dB Spectral MSE Inflation, and Segment-Level Split Leakage:
- Raw unfiltered CHB-MIT pediatric seizure EEG is treated as "clean" ground truth (despite baseline artifacts and a high-frequency floor up to visible in Fig. 6), and computing Spectral MSE after (Eqs. 16–17) artificially inflates baseline errors to whenever wavelet thresholding attenuates high-frequency stopbands near zero. Furthermore, the main synthetic benchmark splits 10-second segments 80/20 across the dataset rather than splitting by patient.
Questions for authors
- Exact Inference Formulation (Clarification): At test time, is computed via on the full signal (as in Eq. 6), or by interleaving and (as in Fig. 1, and at which index sets vs. )? In
iPSD, is re-trained from scratch for epochs on every test segment (taking ), or applied in a single forward pass? - Why Run
lil'UCBAfter Pulling All 35 Deterministic Arms (Clarification)? Since all 35 arms for a test signal are deterministic and Line 1 of Algorithm 2 already evaluates all 35 arms once, why runlil'UCBfor additional pulls of the same deterministic arm instead of picking the best arm after the 35 initial pulls? And why does a single static partition (iPSD-Zero) beat every learned window-varying policy in Table 5 exceptiPSD-GRU? - Selection Bias on WGN Noise Correlation (Clarification): Since fixed interleaved downsampling (
ID) on i.i.d. WGN of length already has true correlation and theoretical finite-sample (matchingID's ), doesn'tiPSD's lower prove that conditions on the noise realization to cancel sample inner products—violating in Appendix C? - Parameter Count & Receptive Field (Clarification): How do you reconcile Section 3's claim of parameters with Table 12's
0.33 m() parameters, and how can a 3-layer Conv1D with a (7-sample) receptive field capture 10-second neural rhythms?
5 comments
[W5] Ground-truth contamination, log-dB S-MSE inflation, and patient-level splitting
Response: Because EEG contamination occurs at acquisition, perfectly artifact-free ground truth cannot be guaranteed. We therefore use high-quality EEG as references, following established evaluation practices. The spectral floor up to 60 Hz in Fig. 6 does not establish strong noise contamination, since clean EEG can contain component at this frequency range. We compute spectral MSE after log so that the metric directly reflects the reconstruction quality of log spectrum, a representation widely used to estimate downstream EEG features, including aperiodic spectral slope and offset. The baselines’ strong attenuations and error on weak spectral components can contribute little to linear-scale spectral MSE, whereas log-spectrum MSE captures these large distortions.
The manuscript already includes leave-one-subject-out evaluation on five CHB-MIT subjects in Appendix K, with the held-out subject and its associated EMG noise unseen during policy training in each fold. The real-world sleep-staging results in Fig. 5 also use leave-one-subject-out evaluation across seven subjects, as clarified in the updated manuscript.
[Q1] Exact inference formulation
Response: We have carefully examined the paper regarding the comments in [W1] and [Q1] and found no contradiction or error in the notation definitions. Equation (2) shows training by predicting the right signal partition from the left and vice versa; but at inference each partition passes through to be denoised and is then placed back in its original indices ( at and at ). Equation (6) describes regularization which uses , i.e. the output of denoising network when the whole is passed through; we never defined as the comment suggests. At test time, the output signal is reconstructed from the two denoised partitions consistent with what is seen in Fig. 1. We have made an extra note around Eq. 6 to avoid any possible misunderstanding. We have also updated the manuscript to make sure that it is clear that, for iPSD, is generalized to the test set along with . Regarding [W1]’s observation that Algorithms 1 and 2 do not explicitly output , we clarify their respective scopes: Algorithm 1 learns the network weights for iPSD, while Algorithm 2 identifies the best partition for iPSD-Zero. In both cases is subsequently reconstructed following the same procedure as mentioned above.
[Q2] Why pulling arms using lil’ UCB? Why does iPSD-Zero beat Table 5 iPSD variants?
Response: As clarified in our response to [W2], each arm pull trains in a non-convex stochastic optimization process, leading to non-deterministic reward. Repeated pulls enable reliable identification of the arm with the highest expected reward. To further support the effectiveness of lil’ UCB-driven arm pulls, we have added an experiment (Appendix L, Table 16, and see results below) comparing iPSD-Zero with two ablated variants that replace lil' UCB with exhaustive and random search. For each test segment, both variants are allocated exactly the number of arm pulls used by iPSD-Zero. Exhaustive search pulls each arm once in a fixed sequential order, repeating this cycle until the budget is exhausted. The results show that, at dB input SNR, iPSD-Zero achieves >1.35 dB higher output SNR under WGN and >1.71 dB higher output SNR under EMG.
| Method | WGN Output SNR (dB) | WGN Output S-MSE | EMG Output SNR (dB) | EMG Output S-MSE |
|---|---|---|---|---|
| Exhaustive search | 2.543 ± 0.839 | 143.996 ± 37.192 | 1.560 ± 1.134 | 162.215 ± 40.662 |
| Random search | 2.541 ± 0.838 | 143.355 ± 37.550 | 1.547 ± 1.121 | 160.687 ± 41.964 |
| iPSD-Zero | 3.894 ± 0.792 | 58.341 ± 25.130 | 3.274 ± 1.140 | 56.620 ± 29.752 |
Regarding the performance of iPSD-Zero, the zero-shot formulation means iPSD-Zero is optimized for every new test segment, while pretrained iPSD generalizes to unseen test segments in a single inference pass without per-segment optimization. The strong performance of iPSD-Zero shows that our zero-shot per-segment optimization (that iPSD does not use) is effective in compensating for the lack of pretraining (that iPSD does use).
[Q3] selection bias on WGN.
Response: We agree that iPSD’s lower reflects the influence of the noise on partitioning, which can violate strict in practice. Together with iPSD’s stronger performance over ID, this supports the method’s practical robustness when the noise assumptions do not hold exactly. To further quantify the effect, we have added an analysis in Appendix I.3 (also shown below) that decomposes iPSD’s observed reward improvement over ID. This analysis shows that the gain is primarily driven by improved clean-signal predictability rather than changes in the noise-related component.
Specifically, the reward in Eq. (4), normalized by signal length , can be written as where is the clean-target prediction error, with lower values indicating greater clean-signal predictability, is the noise-related contribution to the reward, and is the partition-invariant noise energy. The reward improvement of iPSD over ID can therefore be decomposed as:
=R_{\mathrm{iPSD}}-R_{\mathrm{ID}} =\Delta\mathcal L_x+2\Delta C,$$ where $\Delta\mathcal L_x=\mathcal L_{x,\mathrm{ID}}-\mathcal L_{x,\mathrm{iPSD}}$ and $\Delta C=C_{\mathrm{iPSD}}-C_{\mathrm{ID}}$. The results on 200 test segments, each with eight independent WGN realizations at $-5$ dB input SNR, show that 96.8% of iPSD’s reward improvement is driven by lower clean-target prediction error. | | Value [95% CI] | Contribution [95% CI] | |---|:---|---:| | $\Delta\mathcal L_x$ | 0.0646 [0.0627, 0.0664] | 96.8% [95.9, 97.7] | | $2\Delta C$ | 0.0022 [0.0015, 0.0028] | 3.2% [2.3, 4.1] | | $\Delta R$ | 0.0667 [0.0646, 0.0688] | 100% | --- **[Q4]** Different parameter counts reported & small receptive field capturing neural rhythms. **Response**: As clarified in our response to [W4], the denoising network $f_\theta$ has ~7k parameters, while 0.33M counts parameters of the complete iPSD model ($f_\theta$, $\pi_\psi$, and PPO value head). The 10 second neural rhythm is captured by the GRU-based $\pi_\psi$ with the full input as context, whereas the small receptive field (that the reviewer refers to) belongs to the denoising network $f_\theta$ which acts on the partitions from $\pi_\psi$.
[W1] Unspecified Inference Procedure & Contradiction Between Fig. 1, Eq. (2), and Eq. (6).
Response: We carefully checked these components of the manuscript and found no contradiction, and we will explain in our response to [Q1] that the confusion may arise from a misunderstanding of the distinct roles of fθ at inference and in the loss terms used during training. Also, the perceived inconsistency may partly stem from expecting our notation to match a reference (Mansour & Heckel, 2023), but we have tailored and carefully defined the notations which remain consistent within the manuscript.
[W2] Static iPSD-Zero outperforming iPSD candidates; Pointless lil’ UCB exploration over deterministic arms.
Response The comment is disregarding a fundamental principle of zero-shot denoising, i.e., with no prior training the method applies to a new signal seen for the first time. To do this, iPSD-Zero reduces the search space by using a common partitioning across windows, but it optimizes both the partition and the denoiser for every new test segment. Pretrained iPSD instead generalizes to unseen test segments in a single inference pass without per-segment optimization. The zero-shot variant having a performance close to iPSD does not point at a problem here, and it is merely a sign that live optimization on a new input signal using a reduced search space is successful; a specific method design for a specific application.
We also need to state that the non-convex stochastic optimization of fθ is not deterministic, which is the main process driving the non-deterministic reward of iPSD-Zero. The reviewer has missed this, and we have checked the manuscript and made sure this was clear; however, we are happy to improve the clarity if the reviewer can point at the root of this misunderstanding. Please also see our response to [Q2] for an added experiment that further supports the effectiveness of lil’ UCB.
[W3] Selection can induce noise correlation that invalidates Theorem 1 & Appendix C.
Response We would like to state that we do understand that assumption used to prove Theorem 1 may be violated in real data. However, the manuscript approach is standard in reference-free denoising works, starting with analytical results (in Theorem 1 and Appendix C) under explicitly stated assumptions, and motivated by the theory, extensive experiments are performed deliberately involving practical violations of E[nr∣sl,xr]=0 to show the robustness of the proposed method. In our response to this comment and [Q3], we clarify the WGN correlation metric and present a further experiment quantifying the effect of conditional zero-mean violations on the RL reward. Also, thanks for pointing out the miscalculated number in Appendix F which we have now corrected in the updated manuscript.
We would also like to clarify two misunderstandings in the comment. First, iPSD selects index sets while preserving the original temporal order, thus the reviewer’s example of sorting noise is not permitted by our partitioning construction. Second, the subtractions in Theorem 1 involve signals of matching dimensions under the theorem’s definitions.
[W4] Redundancy of using both PSNR and SNR; 7 sample receptive field; parameter discrepancy.
Response The claim on PSNR is based on textbook technical implication of PSNR, so we cannot agree that it is “mathematically false” as the comment asserts. The ~14.11 dB difference between PSNR and SNR means the peak power in the reference EEG is about 26 times the average power, indicating pronounced peaks relative to the overall signal level. Thanks for pointing out the typo which we have now fixed; 18.10 should have been written 18.01. For the receptive field and parameter count, the reported two values refer to different components: the small receptive field and ~7k parameters describe the denoiser fθ, whereas the 0.33M parameters count the full iPSD framework, including fθ, πψ, and the PPO value head. Regarding the concern about the network being unable to span an EEG rhythm, we clarify that our partitioning policy uses GRU, which spans the full 10-second segment to determine the partition.