acceptodds
Under review as a conference paper at ICLR 2027

Enabling Unsupervised Training of Deep EEG Denoisers With Intelligent Partitioning

Abstract

Denoising electroencephalogram (EEG) is an inherently challenging task, since neural activity is not only subtle but also inseparable from spectrally overlapping noise artifacts. Today, effective EEG denoising is more important than ever, given the rapid adoption of wearables across various applications. Deep learning methods have shown promising results in decomposition-free denoising that handles the time-varying pervasive EEG artifacts. However, training highly expressive neural networks requires artifact-free EEG, which is inherently unobtainable. To address this, we propose Intelligent Partitioning for Self-supervised Denoising (iPSD). Our method eliminates the need for clean references by learning to partition an input EEG segment into independent noisy realizations with the same underlying signal. This enables self-supervision of deep learning denoisers, even in zero-shot settings where only a single EEG segment to be denoised is available. We validate iPSD through extensive experiments, including validations on wearable EEG from in-ear sensors. The results show that iPSD achieves state-of-the-art performance, most notably under extremely low signal-to-noise ratios (down to dB) and challenging artifacts (e.g., EMG), with spectral fidelity orders of magnitude higher than competitive baselines.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG denoising, self-supervised learning, wearable biosignals

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Xg52OpenReview ↗
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 s∈RL\mathbf{s} \in \mathbb{R}^L into odd and even samples via fixed interleaved downsampling, iPSD segments s\mathbf{s} into non-overlapping windows of length W=8W = 8 and uses a bidirectional GRU policy network πψ\pi_\psi 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 fθf_\theta 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
  1. 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
  1. Unspecified Inference Procedure & Contradiction Between Figure 1, Eq. (2), and Eq. (6): The paper never defines how the denoised signal x^∈RL\hat{\mathbf{x}} \in \mathbb{R}^L is computed at test time, presenting three mutually incompatible descriptions:
    • Figure 1 (Page 2) shows s∈RL\mathbf{s} \in \mathbb{R}^L split into (sl,sr)∈RL/2×RL/2(\mathbf{s}^l, \mathbf{s}^r) \in \mathbb{R}^{L/2} \times \mathbb{R}^{L/2}, mapped to (fθ(sl),fθ(sr))(f_\theta(\mathbf{s}^l), f_\theta(\mathbf{s}^r)), and interleaved into x^∈RL\hat{\mathbf{x}} \in \mathbb{R}^L. Yet in Eq. (2), fθ(sl)f_\theta(\mathbf{s}^l) is trained to predict sr\mathbf{s}^r (at indices IrI^r) and fθ(sr)f_\theta(\mathbf{s}^r) predicts sl\mathbf{s}^l (at indices IlI^l), leaving unspecified whether fθ(sl)f_\theta(\mathbf{s}^l) is placed at IlI^l or IrI^r.
    • Equation (6) (Page 5) defines the consistency loss using πl(fθ(s))\pi^l(f_\theta(\mathbf{s})) and πr(fθ(s))\pi^r(f_\theta(\mathbf{s})), which applies fθf_\theta directly to the full-length unpartitioned signal s∈RL\mathbf{s} \in \mathbb{R}^L (where Mansour & Heckel, 2023 define x^=fθ(s)\hat{\mathbf{x}} = f_\theta(\mathbf{s})), contradicting Figure 1.
    • Algorithm 1 & Appendix K (Page 24) imply that only the policy πψ\pi_\psi transfers across subjects and fθf_\theta is trained from scratch (∼200\sim 200 epochs) on each test segment in 90 ms90\text{ ms} (Appendix F, line 854), yet neither Algorithm 1 nor Algorithm 2 ever outputs x^\hat{\mathbf{x}}.
  2. The iPSD-Zero Bandit Paradox Refutes the Core Non-Stationarity Premise & Wastes ≥90%\ge 90\% 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, 330k330\text{k} params) to adapt partitions across all 320 local windows. Yet iPSD-Zero uses a single static 8-sample partition repeated identically across all 320 windows—and outperforms every learned window-varying policy in Table 5 except iPSD-GRU (iPSD-MLP, iPSD-RNN, iPSD-U-Net, and DR in Table 2), matches iPSD-GRU within 0.02–0.08 dB0.02\text{--}0.08\text{ dB} on WGN, and produces wearable EEG outputs "visually indistinguishable from those of iPSD" (page 7, line 368).
    • Pointless lil'UCB Exploration Over Deterministic Arms: For a fixed test signal s\mathbf{s} and arm a∈{1,…,35}a \in \{1, \dots, 35\}, the split (sal,sar)(\mathbf{s}_a^l, \mathbf{s}_a^r) is 100% deterministic (the only variance across pulls is random weight initialization of fθf_\theta, σ=0.008\sigma = 0.008, line 840). Algorithm 2 (page 16) already pulls all 35 arms once at initialization (Ta(0)=1T_a(0) = 1), and then runs lil'UCB with β=1,α=[(2+β)/β]2=9\beta = 1, \alpha = [(2+\beta)/\beta]^2 = 9 until Ta⋆(t)≥1+9∑i≠a⋆Ti(t)≥1+9×34=307T_{a^\star}(t) \ge 1 + 9 \sum_{i \neq a^\star} T_i(t) \ge 1 + 9 \times 34 = \mathbf{307} pulls of the winning arm (≥342\ge 342 full 200-epoch trainings of fθf_\theta). Re-training fθf_\theta over 300 times on the exact same deterministic signal pair wastes ≥90%\ge 90\% of compute compared to selecting the best arm after the 35 initial pulls.
  3. 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 E[nr∣sl,xr]=0\mathbb{E}[\mathbf{n}^r \mid \mathbf{s}^l, \mathbf{x}^r] = \mathbf{0}: Because πψ(⋅∣s)\pi_\psi(\cdot \mid \mathbf{s}) observes the entire noisy signal s=x+n\mathbf{s} = \mathbf{x} + \mathbf{n} before choosing (Il,Ir)(I^l, I^r), conditioning on the chosen partition induces selection/collider bias between nl\mathbf{n}^l and nr\mathbf{n}^r: E[nr∣sl,xr]≠0\mathbb{E}[\mathbf{n}^r \mid \mathbf{s}^l, \mathbf{x}^r] \neq \mathbf{0} (e.g., sorting pure noise within each window makes nl\mathbf{n}^l and nr\mathbf{n}^r strongly correlated, minimizing N2N MSE by predicting noise). Moreover, zero MMSE residual E[∥xr−E[xr∣sl]∥22]=0\mathbb{E}[\|\mathbf{x}^r - \mathbb{E}[\mathbf{x}^r \mid \mathbf{s}^l]\|_2^2] = 0 (Eq. 11) does not imply xl≈xr\mathbf{x}^l \approx \mathbf{x}^r (in Appendix B's example xI(k)=(1,−1,1,−1)\mathbf{x}_{I^{(k)}} = (1, -1, 1, -1), xr=−xl\mathbf{x}^r = -\mathbf{x}^l has zero MMSE), Theorem 1 subtracts x∈RL\mathbf{x} \in \mathbb{R}^L from fθ(sl)∈RL/2f_\theta(\mathbf{s}^l) \in \mathbb{R}^{L/2}, and Appendix F (line 863) miscalculates 3532035^{320} as 35×320=11,20035 \times 320 = 11{,}200.
    • WGN Correlation Fallacy (Table 10 & Page 9): Under i.i.d. WGN n∼N(0,σn2IL)\mathbf{n} \sim \mathcal{N}(\mathbf{0}, \sigma_n^2 \mathbf{I}_L) with M=1280M = 1280 samples, fixed interleaved downsampling (ID) already produces 100% independent Gaussian vectors (ρ=0\rho = 0) whose finite-sample ∣r∣×100|r| \times 100 has exact theoretical expectation 2/(π⋅1279)×100=2.23±1.69\sqrt{2 / (\pi \cdot 1279)} \times 100 = \mathbf{2.23 \pm 1.69}—matching ID (2.22±1.552.22 \pm 1.55). Achieving ∣r∣=1.26±0.60<2.23|r| = 1.26 \pm 0.60 < 2.23 on i.i.d. Gaussian noise is not "greater independence"; it proves that πψ\pi_\psi actively selects partitions conditioned on the noise realization n\mathbf{n} to cancel sample inner products, violating noise independence.
  4. Mathematical Redundancy of PSNR with SNR, 7-Sample Receptive Field Paradox, and 45×45\times Parameter Discrepancy:
    • PSNR−SNR≡14.10 dB\text{PSNR} - \text{SNR} \equiv 14.10\text{ dB}: Subtracting Eq. (14) from Eq. (15) cancels x^\hat{\mathbf{x}} identically: PSNR(x,x^)−SNR(x,x^)≡10log⁡10(Lxmax⁡2/∥x∥22)=14.10 dB\text{PSNR}(\mathbf{x}, \hat{\mathbf{x}}) - \text{SNR}(\mathbf{x}, \hat{\mathbf{x}}) \equiv 10 \log_{10}(L x_{\max}^2 / \|\mathbf{x}\|_2^2) = 14.10\text{ dB} across every table. Claiming PSNR "emphasizes robustness to large transient spikes" (line 258) is mathematically false, and Optimal-WT (EMG 0 dB0\text{ dB}) in Tables 1 & 7 has a transcription error (18.10−3.89=14.21 dB≠14.10 dB18.10 - 3.89 = 14.21\text{ dB} \neq 14.10\text{ dB}).
    • Receptive Field & Parameter Contradictions: fθf_\theta (3-layer Conv1D, kernel size 3) has a 7-sample (27.3 ms27.3\text{ ms}) receptive field—unable to span a single cycle of delta/theta/alpha rhythms (contradicting line 268)—and Section 3 claims ∼7k\sim 7\text{k} parameters whereas Table 12 reports 0.33 m (330,000330{,}000) parameters (a 45×45\times discrepancy).
  5. 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 x\mathbf{x} (despite baseline artifacts and a −30 dB/Hz-30\text{ dB/Hz} high-frequency floor up to 60 Hz60\text{ Hz} visible in Fig. 6), and computing Spectral MSE after 10log⁡10(⋅)10\log_{10}(\cdot) (Eqs. 16–17) artificially inflates baseline errors to 103–104 dB210^3\text{--}10^4\text{ dB}^2 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
  1. Exact Inference Formulation (Clarification): At test time, is x^∈RL\hat{\mathbf{x}} \in \mathbb{R}^L computed via x^=fθ(s)\hat{\mathbf{x}} = f_\theta(\mathbf{s}) on the full signal (as in Eq. 6), or by interleaving fθ(sl)f_\theta(\mathbf{s}^l) and fθ(sr)f_\theta(\mathbf{s}^r) (as in Fig. 1, and at which index sets IlI^l vs. IrI^r)? In iPSD, is fθf_\theta re-trained from scratch for ∼200\sim 200 epochs on every test segment (taking 90 ms90\text{ ms}), or applied in a single forward pass?
  2. Why Run lil'UCB After Pulling All 35 Deterministic Arms (Clarification)? Since all 35 arms for a test signal s\mathbf{s} are deterministic and Line 1 of Algorithm 2 already evaluates all 35 arms once, why run lil'UCB for ≥307\ge 307 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 except iPSD-GRU?
  3. Selection Bias on WGN Noise Correlation (Clarification): Since fixed interleaved downsampling (ID) on i.i.d. WGN of length M=1280M = 1280 already has true correlation ρ=0\rho = 0 and theoretical finite-sample E[∣r∣]×100=2.23±1.69\mathbb{E}[|r|] \times 100 = 2.23 \pm 1.69 (matching ID's 2.22±1.552.22 \pm 1.55), doesn't iPSD's lower ∣r∣=1.26±0.60|r| = 1.26 \pm 0.60 prove that πψ\pi_\psi conditions on the noise realization n\mathbf{n} to cancel sample inner products—violating E[nr∣sl,xr]=0\mathbb{E}[\mathbf{n}^r \mid \mathbf{s}^l, \mathbf{x}^r] = \mathbf{0} in Appendix C?
  4. Parameter Count & Receptive Field (Clarification): How do you reconcile Section 3's claim of ∼7k\sim 7\text{k} parameters with Table 12's 0.33 m (330k330\text{k}) parameters, and how can a 3-layer Conv1D with a 27.3 ms27.3\text{ ms} (7-sample) receptive field capture 10-second neural rhythms?
5 comments

[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 fθf_\theta 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 fθf_\theta 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\mathbb E[n^r\mid s^l,x^r]=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; 18.1018.10 should have been written 18.0118.01. 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 fθf_\theta, whereas the 0.33M parameters count the full iPSD framework, including fθf_\theta, πψ\pi_\psi, 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.

[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 srs^r from the left fθ(sl)f_\theta(s^l) and vice versa; but at inference each partition passes through fθf_\theta to be denoised and is then placed back in its original indices (fθ(sl)f_\theta(s^l) at IlI^l and fθ(sr)f_\theta(s^r) at IrI^r). Equation (6) describes regularization which uses fθ(s)f_\theta(s), i.e. the output of denoising network when the whole ss is passed through; we never defined x^=fθ(s)\hat{x}=f_\theta(s) as the comment suggests. At test time, the output signal x^\hat{x} 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, fθf_\theta is generalized to the test set along with πψ\pi_\psi. Regarding [W1]’s observation that Algorithms 1 and 2 do not explicitly output x^\hat{x}, 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 x^\hat{x} 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 fθf_\theta 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 −5-5 dB input SNR, iPSD-Zero achieves >1.35 dB higher output SNR under WGN and >1.71 dB higher output SNR under EMG.

MethodWGN Output SNR (dB)WGN Output S-MSEEMG Output SNR (dB)EMG Output S-MSE
Exhaustive search2.543 ± 0.839143.996 ± 37.1921.560 ± 1.134162.215 ± 40.662
Random search2.541 ± 0.838143.355 ± 37.5501.547 ± 1.121160.687 ± 41.964
iPSD-Zero3.894 ± 0.79258.341 ± 25.1303.274 ± 1.14056.620 ± 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 ∣r∣|r| reflects the influence of the noise on partitioning, which can violate strict E[nr∣sl,xr]=0\mathbb E[n^r\mid s^l,x^r]=0 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 LL, can be written as R=−Lx+2C−k,R=-\mathcal L_x+2C-k, where Lx=1LE[∥fθ(sl)−xr∥2+∥fθ(sr)−xl∥2]\mathcal L_x=\frac{1}{L}\mathbb E[\|f_\theta(s^l)-x^r\|^2+\|f_\theta(s^r)-x^l\|^2] is the clean-target prediction error, with lower values indicating greater clean-signal predictability, C=1LE[(nr)⊤(fθ(sl)−xr)+(nl)⊤(fθ(sr)−xl)]C=\frac{1}{L}\mathbb E[(n^r)^\top(f_\theta(s^l)-x^r)+(n^l)^\top(f_\theta(s^r)-x^l)] is the noise-related contribution to the reward, and k=1LE∥n∥2k=\frac{1}{L}\mathbb E\|n\|^2 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 $\pi_\psi$.
Reviewer_1rCSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Intelligent Partitioning for Self-supervised Denoising (iPSD), a framework for EEG denoising without clean reference signals. The method combines Noise2Noise-based self-supervised learning with reinforcement learning to optimize the partitioning of noisy EEG signals. The authors further introduce iPSD-Zero, which employs a multi-armed bandit algorithm to identify suitable partitions without prior training. Experiments on synthetic noise contamination and real-world wearable EEG demonstrate improvements in denoising performance and downstream sleep-stage classification.

Critical strengths

The paper addresses an important practical problem in EEG signal processing, particularly the difficulty of obtaining clean EEG references for supervised denoising.

The proposed framework achieves promising experimental results under challenging noise conditions. The evaluation on real-world in-ear EEG also demonstrates its potential for practical applications.

Critical weaknesses

Limited methodological novelty. The proposed method mainly combines existing techniques, including Noise2Noise, reinforcement learning (PPO), and multi-armed bandit optimization. The central idea of constructing noisy observation pairs for self-supervised denoising has already been explored in previous work, particularly Zero-Shot Noise2Noise. Replacing fixed partitioning with an RL-based partitioning strategy is primarily an extension of this existing framework rather than a fundamentally new learning paradigm. Overall, the contribution appears more engineering- and application-oriented, and the authors should more clearly articulate the methodological novelty and fundamental insights provided to the machine learning community.

Insufficient clarity in the reinforcement learning formulation. Although the paper provides mathematical definitions of the partitioning strategy and reward, the RL formulation is not presented clearly and systematically. In particular, the authors should explicitly describe the state representation, action space, reward function, and policy optimization procedure, and explain how they interact during training.

Insufficient evaluation across different EEG artifact types and noise levels. The experiments mainly focus on white Gaussian noise and EMG contamination. Although the authors evaluate EMG under several SNR conditions, EOG artifacts, which are another major source of EEG contamination, are not sufficiently investigated. The authors should conduct systematic experiments on both EMG and EOG artifacts across a consistent range of input SNRs.

Questions for authors

What are the fundamental methodological contributions of iPSD beyond combining existing Noise2Noise and reinforcement learning techniques? How does the proposed approach advance self-supervised learning beyond its application to EEG denoising?

Could the authors provide a clearer and more systematic formulation of the reinforcement learning process, including its states, actions, rewards, and policy optimization? What evidence demonstrates that the proposed reward is aligned with actual EEG denoising quality?

Can the authors extend their experiments to include EOG and mixed EMG–EOG artifacts under different SNR conditions?

Why does the method partition EEG signals into exactly two sub-signals? Have the authors investigated three-view or multi-view alternatives, and how would the number of partitions affect denoising performance and computational efficiency?

2 comments

[W1]: Articulation of methodological novelty and insights for the machine learning community.

Response: The core methodological contribution of our proposed method is the RL reward formulation that extends ground-truth-free training from learning the denoiser to also learning adaptive partitions, enabling the construction of stronger self-supervision. We acknowledge the application-oriented engineering noted by the reviewer and have now highlighted this core methodological contribution more clearly in the updated manuscript. Please also see our response to [Q1].


[W2]: Missing systematic presentation of the RL formulation.

Response: We have directly resolved this in the updated manuscript. We have added Table 19 (also inserted below), which explicitly summarizes the definition of each RL component and explains their interactions during training.

ComponentDefinition in iPSD
State ssThe noisy EEG segment s=x+n∈RLs=x+n\in\mathbb R^L.
Action aaA partition a=(Il,Ir)a=(I^l,I^r), where the index subsets are disjoint, cover {1,…,L}\{1,\ldots,L\}, and each contain L/2L/2 indices.
Reward RRThe negative converged self-supervised N2N loss value, as defined in Eq. (4).
Training interactionSample ss; generate (sl,sr)(s^l,s^r) using a∼πψ(⋅∣s)a\sim\pi_\psi(\cdot\mid s); update fθf_\theta to convergence; compute RR; update πψ\pi_\psi using PPO based on Eq. (5); repeat (Detailed explanation can be found in Algorithm 1.).

[W3]: Investigation of other noise types such as EOG and mixed EMG–EOG.

Response: We have added experiments on both EOG and mixed EMG–EOG artifacts to Appendix N following the eye-blink contamination construction in [Sci Rep. 2021; 11:11043]. The results show that iPSD consistently outperforms all baseline methods under EOG and mixed EMG–EOG contamination. See our response to [Q3], where we provide more detail and discussions on the experiment.


[Q1] What are the fundamental methodological contributions of iPSD beyond combining existing Noise2Noise and reinforcement learning techniques? How does the proposed approach advance self-supervised learning beyond its application to EEG denoising?

Response: The absence of any ground truth is the crux of EEG denoising naturally, although it can be the challenge in a variety of other tasks. Our proposed formulation of the RL reward extends ground-truth-free training from learning the denoiser to also learning adaptive partitions. This extension enables the construction of stronger self-supervision from noisy observations alone, which not only contributes to the broader problem of reference-free denoising, but also benefits settings with limited clean supervision, for example by supporting pretraining on abundant noisy data before fine-tuning.


[Q2] Could the authors provide a clearer and more systematic formulation of the reinforcement learning process, including its states, actions, rewards, and policy optimization? What evidence demonstrates that the proposed reward is aligned with actual EEG denoising quality?

Response: Yes, absolutely; we have updated the manuscript with inclusion of a systematic RL formulation in Table 19. There are two pieces of evidence on reward alignment. One being the strong performance against Interleaved Downsampling (ID), which is not optimized with the proposed reward. Following this comment, we add another in Appendix O, Figure 7 that provides more direct evidence. For synthetic signals with WGN at -5 dB, we unpack instances during the training of iPSD, visualizing true denoised SNR versus reward in every instance. The results show strong correlation ( r>0.98r>0.98) between reward and denoising quality measured by denoised SNR.

[Q3] Can the authors extend their experiments to include EOG and mixed EMG–EOG artifacts under different SNR conditions?

Response: In response to this comment, we have added experiments with EOG and mixed EMG-EOG artifacts to the updated manuscript (see Appendix N). For EOG, we follow the eye-blink contamination construction in [Sci Rep. 2021; 11:11043]. For mixed EMG–EOG contamination, we combine each blink artifact segments with an EMG artifact used in Table 1 at equal power. For input SNRs of −5-5 and 00 dB, we report SNR gain and post-denoising S-MSE (as seen in the tables below). The results show that iPSD consistently outperforms all baseline methods under EOG and mixed EMG–EOG contamination. Overall, gains are smaller under these artifacts compared to WGN and EMG, highlighting the difficulty of suppressing ocular artifacts without distortion using a single EEG channel without EOG reference. We now discuss this in the updated manuscript, as well as multichannel extensions for improved EOG–EEG separation as future work.

EOG:

MethodSNR Gain (−5 dB input)S-MSE (−5 dB input)SNR Gain (0 dB input)S-MSE (0 dB input)
mVMD−3.775.66×10⁴-4.184.31×10⁴
CEEMDAN−1.982.47×10⁴-2.132.08×10⁴
FrWT−2.623.39×10⁴-3.062.94×10⁴
EMD-LoG+0.4560.53×10⁴+0.3940.43×10⁴
WPT-ICA+1.3860.31×10⁴+0.6610.23×10⁴
Chen et al.-0.1530.32×10⁴-0.5060.24×10⁴
Optimal-WT+1.5440.24×10⁴+0.7110.22×10⁴
iPSD+3.07292.65+2.82553.63
iPSD-Zero+2.744102.45+2.17084.46

EMG-EOG Mixed:

MethodSNR Gain (−5 dB input)S-MSE (−5 dB input)SNR Gain (0 dB input)S-MSE (0 dB input)
mVMD−2.8525.46×10⁴-2.7463.93×10⁴
CEEMDAN-0.0811.27×10⁴-0.5221.02×10⁴
FrWT−0.6192.39×10⁴-1.4162.24×10⁴
EMD-LoG+2.0480.45×10⁴+1.5690.35×10⁴
WPT-ICA+2.5590.27×10⁴+1.2430.19×10⁴
Chen et al.+1.690.04×10⁴+0.910.02×10⁴
Optimal-WT+2.8530.03×10⁴+1.6090.02×10⁴
iPSD+5.49561.27+4.36348.91
iPSD-Zero+4.95269.35+3.61854.17

[Q4] Why does the method partition EEG signals into exactly two sub-signals? Have the authors investigated three-view or multi-view alternatives, and how would the number of partitions affect denoising performance and computational efficiency?

Response: We have expanded the discussion in Appendix F to justify our two-view design on stronger grounds. First, using more views would mean retaining fewer samples in each view, leading to more aggressive subsampling that compromises clean-component similarity. Second, multi-view partitioning can drastically expand the action space: our two-view design with eight-sample windows admits 35 candidates per window, compared with 280 for three views with nine-sample windows and over two million for four views with 16-sample windows. This increases the number of policy parameters and makes policy learning less tractable. We do have results from three-view alternative using the same policy backbone as iPSD and observed a slight drop in performance. We use nine-sample local windows, padding inputs to a multiple of nine and cropping reconstructed outputs back to the original length. The policy output layer is expanded to 280 outputs per window to accommodate the candidate partitions. Training uses all six directed cross-view N2N loss terms. At input SNRs of −5-5 and 00 dB, the three-view variant achieves output SNRs of 3.71 and 5.72 dB for WGN, and 3.54 and 5.87 dB for EMG, respectively.

Reviewer_kybSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes iPSD, which uses reinforcement learning to partition a single noisy EEG recording into two sub-signals and constructs Nosie2Noise-style self-supervised training pairs. It also provides a zero-shot variant, iPSD-Zero. The experiments cover synthetically noise-added EEG and sleep staging on real in-ear EEG.

Critical strengths

Given the practical difficulty of lacking clean EEG references, the learnable partitioning strategy is well motivated; iPSD-Zero directly optimizes on a single noisy signal and has potential practical deployment value. The experiments on synthetically noise-added data are generally strong: iPSD outperforms various conventional and self-supervised baselines in SNR, PSNR, and Spectral MSE. The appendix provides relatively extensive supplementary experiments, including partitioning ablations, robustness to correlated noise, cross-subject validation, and low-sampling-rate tests, which strengthen the completeness of the paper.

Critical weaknesses

The theory does not yet cover the actual learned adaptive partitioning mechanism. Theorem 1 and Appendix C rely on noise independence and conditional zero-mean assumptions, but the partitioning policy π_ψselects samples based on the noisy observation and may violate these conditions. The signal-mismatch/noise-correlation results in Table 10 and the AR-noise experiments in Appendix I.6 provide only indirect empirical support; low Pearson correlation cannot prove independence or conditional zero mean. The order-of-magnitude gain is not unique to intelligent partitioning, and its sources are not sufficiently separated. Table 2 shows that fixed interleaved downsampling (ID) already reduces Spectral MSE from the thousands/hundreds scale to the tens scale, although iPSD still gives consistent additional gains over ID. The current experiments do not separate the contributions of the denoising network, regularization, self-supervised framework, and RL partitioning, nor do they compare against random or exhaustive partition search under the same total computational budget. The authors should clarify these contributions and add an efficiency comparison. Downstream classification performance on real in-ear EEG does not establish clinical-grade waveform fidelity. The sleep-staging results support that iPSD preserves downstream-useful information, but the main text does not specify whether the evaluation is subject-independent or report a raw noisy-EEG baseline. Training a separate classifier per denoising method may reflect adaptation to method-specific distortions and cannot directly prove waveform fidelity or rule out over-smoothing. Please add a subject-independent evaluation, a raw noisy baseline, and quantitative metrics for key waveforms and band power, and soften claims that infer overall signal quality from classification performance.

Questions for authors

Can the authors further examine the noise conditions after adaptive partitioning on synthetic data, for example by estimating E[n^rⓜ,∣s^l x^r ]or performing nonlinear conditional dependence tests? Can they decompose the reward signal to show that its improvement mainly comes from the predictability of the clean component rather than from predictable noise?

Under the same total computational budget, can random search or exhaustive search over candidate partitions achieve performance close to iPSD? How much additional benefit does the learned partitioning strategy provide over these simple search methods?

Does the sleep-staging evaluation on real in-ear EEG use a subject-independent split? Can the authors add a raw noisy EEG baseline, a unified-classifier comparison, and quantitative validation of the preservation of key sleep waveforms and band power, so as to more directly support denoising fidelity?

2 comments

[W1] Noise independence and conditional zero-mean in practical setting.

Response: We follow the standard approach of deriving analytical results under explicitly stated assumptions, followed by extensive experiments to show the robustness of the proposed method to assumption violations in practical settings. Noting this comment, we have obtained further results that examine the violation of conditional zero mean (E[nr∣sl,xr]=0\mathbb E[n^r\mid s^l,x^r]=0) by estimating the noise–error cross-term: C=1LE[(nr)⊤(fθ(sl)−xr)+(nl)⊤(fθ(sr)−xl)].C=\frac{1}{L}\mathbb E[(n^r)^\top(f_\theta(s^l)-x^r)+(n^l)^\top(f_\theta(s^r)-x^l)]. Details are provided in our response to [Q1].


[W2] Insufficient results on separate contributions of method components.

Response: Before we describe the actions taken to address this, first note that the core contribution of iPSD is “adaptive partitioning,” whose effect is isolated by the comparison with ID in Table 2. (The ID baseline is an iPSD parallel only with fixed interleaved partitioning while keeping the denoising network, regularization, and self-supervised loss unchanged.) Second, the existing results in Appendix I.5 demonstrate the impact of denoising network via comparisons between different architectures; we have now highlighted this better in the manuscript. Following the reviewer’s comment, to better separate the sources of performance gain we have added comparisons against random and exhaustive search under the same search budget in Appendix L, Table 16. For details, please see our response to [Q2].


[W3] Missing raw noisy-EEG baseline for sleep staging and possible adaptation of classifier to denoising method.

Response: In response to the comment, we have added the raw noisy-EEG baseline to Fig. 5, and have clarified in the updated manuscript that the sleep-staging results used leave-one-subject-out cross-validation across seven subjects. Compared to raw noisy EEG, iPSD improves classification accuracy from 57.1% to 87.0% and κ\kappa from 0.427 to 0.827. To address potential adaptation to method-specific distortions, we additionally report a unified-classifier comparison in Appendix M, Table 17, showing that iPSD’s gains persist without method-specific classifier training. Details on this added comparison are in our response to [Q3]. Lastly, we would like to clarify that our claim from sleep stage classification results in the original manuscript (lines 423–424) only concerns the recovery of meaningful information for sleep-staging, rather than overall waveform fidelity. Because clean references are unavailable for these real-world recordings, quantitative waveform-fidelity metrics are evaluated in the semi-synthetic experiments.


[Q1] Can the authors further examine the noise conditions after adaptive partitioning on synthetic data, for example by estimating E[nr,∣slxr]E[n^r,∣s^l x^r ] or performing nonlinear conditional dependence tests? Can they decompose the reward signal to show that its improvement mainly comes from the predictability of the clean component rather than from predictable noise?

Response: Following this comment, we have added an experiment (see Appendix I.3) to quantify practical violations of E[nr∣sl,xr]=0\mathbb E[n^r\mid s^l,x^r]=0 by estimating C=1LE[(nr)⊤(fθ(sl)−xr)+(nl)⊤(fθ(sr)−xl)]C=\frac{1}{L}\mathbb E[(n^r)^\top(f_\theta(s^l)-x^r)+(n^l)^\top(f_\theta(s^r)-x^l)]. We have also decomposed the reward signal, which shows that iPSD’s reward optimization is primarily driven by improved clean-signal predictability rather than changes in the noise-related component. The details are provided below:

The reward in Eq. (4), normalized by signal length LL, can be written as R=−Lx+2C−k,R=-\mathcal L_x+2C-k, where Lx=1LE[∥fθ(sl)−xr∥2+∥fθ(sr)−xl∥2]\mathcal L_x=\frac{1}{L}\mathbb E[\|f_\theta(s^l)-x^r\|^2+\|f_\theta(s^r)-x^l\|^2] is the clean-target prediction error, with lower values indicating greater clean-signal predictability, C=1LE[(nr)⊤(fθ(sl)−xr)+(nl)⊤(fθ(sr)−xl)]C=\frac{1}{L}\mathbb E[(n^r)^\top(f_\theta(s^l)-x^r)+(n^l)^\top(f_\theta(s^r)-x^l)] is the noise-related contribution to the reward, and k=1LE∥n∥2k=\frac{1}{L}\mathbb E\|n\|^2 is the partition-invariant noise energy. We compare iPSD reward with ID with a static partition:

=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 result (see Appendix I.3 and below) on 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% |

[Q2] Under the same total computational budget, can random search or exhaustive search over candidate partitions achieve performance close to iPSD? How much additional benefit does the learned partitioning strategy provide over these simple search methods?

Response: In the updated manuscript, we have added new comparisons against random and exhaustive search for the zero-shot variant (see Table 16 in Appendix L, also provided below). Each baseline is given exactly the number of searches used by iPSD-Zero on the same test segment. At −5-5 dB input SNR, iPSD-Zero exceeds both baselines by >1.35 dB under WGN and >1.71 dB under EMG, corresponding to >26.7% and >32.6% lower noise power, respectively. For iPSD, this comparison is not directly applicable, as it denoises each test segment in a single forward pass without per-segment search budget.

MethodWGN Output SNR (dB)WGN Output S-MSEEMG Output SNR (dB)EMG Output S-MSE
Exhaustive search2.543 ± 0.839143.996 ± 37.1921.560 ± 1.134162.215 ± 40.662
Random search2.541 ± 0.838143.355 ± 37.5501.547 ± 1.121160.687 ± 41.964
iPSD-Zero3.894 ± 0.79258.341 ± 25.1303.274 ± 1.14056.620 ± 29.752

[Q3] Does the sleep-staging evaluation on real in-ear EEG use a subject-independent split? Can the authors add a raw noisy EEG baseline, a unified-classifier comparison, and quantitative validation of the preservation of key sleep waveforms and band power, so as to more directly support denoising fidelity?

Response: We thank the reviewer for the questions and have addressed them in the updated manuscript; please also see the response to [W3]. The results of Fig. 5 are obtained using leave-one-subject-out cross-validation across seven subjects; we made sure this is clear in the text. We have added raw noisy-EEG baseline to Fig. 5 following the comment. To more directly support denoising fidelity, we have added a unified classifier comparison, now in Appendix M Table 17 and shown below. In each leave-one-subject-out fold, we trained the classifier on scalp EEG from the training subjects and then apply it to the held-out subject’s raw, Optimal-WT-denoised, and iPSD-denoised in-ear EEG. Using the same trained classifier, iPSD-denoised EEG results in 86.81% accuracy, beating both raw in-ear EEG (at 56.41%) and Optimal-WT (at 79.12%), and approaching the scalp-EEG accuracy of 88.07%. The new results from the unified classifier comparisons confirm the iPSD’s gains do not depend on classifier adaptation to method-specific distortions.

MethodAccuracyκ\kappaPrecisionRecall
Raw56.41%0.41861.85%49.48%
Optimal-WT79.12%0.72470.20%70.20%
iPSD86.81%0.82581.20%78.95%
Scalp88.07%0.84183.02%80.05%

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.