DenoFlow: Flow Matching for SSVEP Denoising under Real Physiological Artifacts

arXiv:2610.08817 · cs.LG · Submitted 2026-09-24 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "DenoFlow: Flow Matching for SSVEP Denoising under Real Physiological Artifacts".

Jane: The gist: DenoFlow proposes casting SSVEP denoising as transport by regressing the velocity of the straight path between a contaminated trial and its clean counterpart,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So, we're looking at this paper called "DenoFlow: Flow Matching for SSVEP Denoising under Real Physiological Artifacts." Jane, what do you make of that title? It sounds pretty technical.

Jane: It definitely sounds technical, Tom. It points straight to the problem they are tackling which is SSVEP denoising, but it also specifies two really important things: flow matching and real physiological artifacts.

Tom: Flow matching is this new way they are looking at the problem instead of just using older methods like diffusion models or GANs. And real physiological artifacts means they aren't just dealing with clean noise; they’re dealing with stuff that actually happens in a lab or in a real-world setting, like muscle twitches or eye movements.

Lu: I think the core idea is moving away from traditional mappings and toward something more dynamic, which flow matching seems to enable. It suggests learning the path between a noisy signal and its clean version, rather than trying to learn one single direct transformation from the bad input to the good output.

Meng: If they are dealing with real artifacts, that immediately makes me wonder about how much of this actually transfers when you try to deploy it. Can this model handle all those different types of noise consistently?

Lalam: My analysis shows that their approach is specifically designed to address the issues where models tend to just reproduce what they see instead of cleaning the signal, which is a common problem in these kinds of denoising tasks.

The paper's summary: Tom: Okay, so what's the actual mechanism they are proposing in "DenoFlow: Flow Matching for SSVEP Denoising under Real Physiological Artifacts"? Jane, can you lay out the core idea in plain English?

Jane: They propose treating SSVEP denoising as a transport problem. Instead of learning a direct map from a noisy trial to its clean counterpart, they have a field network that regresses the velocity of the straight path between those two points. Think of it like drawing a smooth trajectory through the noise to get to the clean signal.

Tom: So instead of just trying to get the final clean signal, they are learning *how* to move from noisy data to clean data step by step, using this field velocity as their main training target.

Lu: Exactly. This field network acts like an encoder-decoder where the noisy trial is used as a condition throughout every layer, but it only sees the path information once at a Transformer bottleneck, which lets it describe "the trial as a whole rather than any part of it" #pg8. That's a key representation learning idea here.

Meng: But what about the noise? They mention integrating this learned field forward from the observation using Euler steps, like X hat at step plus one equals X hat at step plus one plus one over S f theta (X hat at step, X double prime, and s over S) #pg6. That sounds like a way to actually reconstruct the signal after training is done.

Lalam: That integration part is crucial because it means we aren't just getting a denoised image; we are doing a process of forward integration using the learned velocity field to get the final result, which gives them more control than just generating an output from scratch.

The paper's improvements: Tom: So where does DenoFlow actually beat what we currently have? What specific improvements are they highlighting in "DenoFlow: Flow Matching for SSVEP Denoising under Real Physiological Artifacts"?

Jane: One big thing they emphasize is that their objective isn't just waveform distance. They use three losses together, and the classification loss makes sure the signal still has that stimulus-related structure preserved, so it doesn't just smooth everything away.

Tom: That means they are explicitly supervising the output to stay decodable, which is a big deal because if you only optimize for waveform distance, you can end up suppressing both the noise and the actual brain response.

Lu: They also have this flow matching velocity regression loss, Lflow, which supervises the network at every point of that path t t-X double prime - X double minus X prime, meaning they are regularizing the entire trajectory, not just focusing on a single endpoint #pg7.

Meng: And then there's the reconstruction loss, Lrec. They use that to constrain where the final output lands at the end of that path, which helps ensure that whatever downstream decoder you use gets something meaningful to decode #pg7.

Lalam: The ablation study showed that removing the reconstruction loss actually caused a big drop in accuracy by three point two five percent, so it proves that constraining the endpoint is a really important piece of supervision for this method <ref:2610.08817#pg2>.

Conclusion: Tom: So to wrap up on "DenoFlow: Flow Matching for SSVEP Denoising under Real Physiological Artifacts," what’s the big picture here? What does it mean when they say they outperformed seven baseline denoising models on both signal fidelity and downstream decoding accuracy?

Jane: It means that this approach handles those real physiological artifacts much better than the previous methods. For instance, under EOG artifact, DenoFlow was near zero there and was the only method to raise that ratio on Benchmark #pg10.

Lu: The implication is that casting denoising as transport allows you to learn a much more robust representation of the underlying signal structure, rather than just fitting a statistical model to the noise.

Meng: From an engineering standpoint, it means training is simplified because they avoid the complexity of adversarial games and they have these complementary losses that supervise everything nicely without needing massive amounts of perfectly paired clean data.

Lalam: And for us using this paper, it’s a strong signal that we can use this path-based learning idea to tackle other kinds of noise in EEG systems, provided we can manage the setup where you manually inject those known perturbations, which they did by injecting EMG and EOG signals under different conditions #pg11.

Tom: So yeah, DenoFlow is a method that reduces training to regression onto the velocity target without needing an adversarial min-max game. It's a solid result against seven baseline models in terms of both how clean the signal looks and how well it decodes after denoising.

Jane: It’s a lot of practical work, especially since they show that performance actually increases as the artifact degradation gets worse, which tells us this method is resilient under more challenging conditions #pg12.

Lu: They also showed that the reconstruction loss made the largest contribution to their final score, which shows that constraining the endpoint is a really important piece of supervision for this method.

Meng: I just wonder about deployment scale; running this forward integration might be computationally heavier than a simple single pass denoiser, but the stable plateau they found with S greater than or equal to two suggests it's manageable #pg8.

Lalam: And overall, "DenoFlow: Flow Matching for SSVEP Denoising under Real Physiological Artifacts" casts denoising as transport by regressing the velocity of the straight path and integrating this field forward from the observation #pg2.

Zhentao Hea, Ziwei Wanga, Dongrui Wua

Hubei Key Laboratory of Brain inspired Intelligent Systems · School of Artificial Intelligence and Automation, Huazhong University of Science and Technology

cs.LG

Submitted: 2026-09-24

Updated: 2026-09-24

Comments: 14 pages, 6 figures

Code: https://github.com/wzwvv/DenoFlow

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: The gist: DenoFlow proposes casting SSVEP denoising as transport by regressing the velocity of the straight path between a contaminated trial and its clean counterpart, integrating this field forward

Key concepts

Transport Formulation: This approach treats denoising as moving from a contaminated signal to its clean counterpart along a learned path. Instead of trying to learn the entire transformation at once, it focuses on learning how fast you need to move along that specific path, which makes the training process more stable and robust.
Field Network: This is an encoder-decoder network that processes EEG data at every layer. It takes the contaminated trial as input and learns to predict the velocity of the straight line path connecting it to a clean version of itself. This network captures complex signal features across multiple stages.
Integration by Forward Propagation: Once trained, this learned 'field' is used to denoise new, unseen data. The process involves integrating this field forward step-by-step from the initial observation using Euler steps. This allows the model to generate a clean output based on the learned trajectory.
Rectified Flow Formulation: This mathematical framework defines how the velocity of the path is calculated in this transport problem. It ensures that the learned field accurately represents a consistent direction and speed for moving from noise toward a clean state, guiding the denoising process effectively.

Terminology

Flow Matching — 1× in this paper · explained in 17 episodes

Summary

The gist: DenoFlow proposes casting SSVEP denoising as transport by regressing the velocity of the straight path between a contaminated trial and its clean counterpart, integrating this field forward from the observation, and supervising it with a classifier trained jointly with the field network.

Problem Addressed

Electroencephalography (EEG)-based brain-computer interfaces (BCIs), particularly steady-state visual evoked potential (SSVEP) systems, are highly vulnerable to noise and artifacts, which severely degrade decoding accuracy <ref:2610.08817#pg2> The paper highlights several challenges in EEG denoising, including traditional methods resting on strong assumptions about noise independence and signal stationarity <ref:2610.08817#pg3>, the lack of clean counterparts for supervised training on public datasets <ref:2610.08817#pg4>, models settling on reproducing their input, and objectives defined solely on waveform distance which fails to constrain the downstream decoder output <ref:2610.08817#pg5>.

Proposed DenoFlow Mechanism

DenoFlow treats SSVEP denoising as learning the path from a contaminated trial to its clean counterpart rather than a direct mapping between them <ref:2610.08817#pg6>. The core of the approach involves two main components:

Field Network: This network is trained to regress the velocity of the straight path between them, following the rectified-flow formulation <ref:2610.08817#pg7>. It acts as an encoder-decoder where the contaminated trial is regarded as a conditioning plane through every layer, while the path position enters once at a Transformer bottleneck, describing the trial as a whole rather than any part of it <ref:2610.08817#pg8>.
Denoising by Integration: After training, denoising integrates this learned field forward from the observation using Euler steps <ref:2610.08817#pg6>. The process is defined by the recursion Xˆ(s+1) = Xˆ(s) + 1/S fθ(Xˆ(s), X˜′, s/S) by eq. (5), where S tr steps are used for training and S steps for inference <ref:2610.08817#pg6>.

Training Objective

The overall training objective combines three losses to supervise the network:

** Lcls = CE(C(Xˆ), y) is the classification loss, which ensures stimulus-related structure is preserved rather than merely smoothed away <ref:2610.08817#pg9>. The classifier C is trained alongside the field network on both reference and integrated outputs <ref:2610.08817#pg9>. **

** Lflow = Et∼U(0,1) 1/NcT fθ xt, X˜′, t − X˜ − X′ is the flow matching velocity regression loss, where t is resampled at each step to supervise the network at every point of the path <ref:2610.08817#pg7>. **

** Lrec = 1/NcT Xˆ − X˜ 2F is the reconstruction loss, which constrains the endpoint that a downstream decoder receives <ref:2610.08817#pg7>. **

The overall training objective is L = Lcls + λflowLflow + λrecLrec, where both trade-off weights are set to one unless stated otherwise <ref:2610.08817#pg8>.

Physiological Artifact Injection and Experimental Setup

To obtain paired noised training data on datasets without ground truth, the authors manually inject the known perturbations to EEG recordings <ref:2610.08817#pg10>. Physiological artifacts are injected from recorded EMG and EOG libraries, considering three conditions: EMG only, EOG only, and EMG-EOG mixed <ref:2610.08817#pg11>. The contamination severity is controlled by the mask fraction and target ratio γ <ref:2610.08817#pg13>. For instance, in the mixed condition, the mixed condition draws one from each library and is therefore some 3 dB harsher <ref:2610.08817#pg13>.

Experimental Results

Experiments were conducted on two public SSVEP datasets with five popular SSVEP decoders <ref:2610.08817#pg2> DenoFlow outperformed seven baseline denoising models on both signal fidelity and downstream decoding accuracy <ref:2610.08817#pg6>. For example, under EOG artifact, DenoFlow was near zero there and is the only method to raise the ratio on Benchmark <ref:2610.08817#pg10>. The results showed that DenoFlow achieved the highest average accuracy across all artifact types on both datasets <ref:2610.08817#pg11>. Furthermore, the benefit of denoising generally increased with the severity of artifact-induced degradation <ref:2610.08817#pg12> and DenoFlow reduced the sensitivity of decoding performance to contamination severity within the evaluated range.

Ablation Study

The ablation study showed that the reconstruction loss made the largest contribution, as removing it reduced accuracy by 3.25% and degraded all three fidelity metrics. Similarly, removing the classification and flow-matching losses reduced accuracy by 1.86% and 1.70%, respectively. The complete objective achieved the best result for all four metrics, indicating that the three loss terms provide complementary supervision.

Parameter Sensitivity

Sensitivity analysis showed that performance was relatively insensitive to moderate changes in any individual loss weight within the evaluated range, with the maximum accuracy difference being only 0.67%. Additionally, using a single inference step (S = 1) consistently produced lower accuracy across tested values of S tr, while settings with S ≥ 2 formed a stable plateau. The default configuration (S tr = 4, S = 4) lies within this stable region.

Conclusion

DenoFlow successfully casts SSVEP denoising as transport by regressing the velocity of the straight path and integrating this field forward from the observation <ref:2610.08817#pg2>. The method reduces training to regression without an adversarial min-max game, and experiments demonstrated that DenoFlow outperformed seven baseline models on both signal fidelity and downstream decoding accuracy <ref:2610.08817#pg2>. Future work will explore extensions to online and cross-subject SSVEP decoding scenarios, as well as lightweight denoising model architectures for real-time deployment <ref:2610.08817#pg2>.

References

[1] Rosenfeld, J.V., Wong, Y.T., 2017. Neurobionics and the brain-computer interface: Current applications and future horizons. Medical Journal of Australia 206, 363–368 <ref:2610.08817#pg2>

[2] Wu, D., Xu, Y., Lu, B.L., 2020. Transfer learning for EEG-based brain-computer interfaces: A review of progress made since 2016. IEEE Transactions on Cognitive and Developmental Systems 14, 4–19 <ref:2610.08817#pg3>

[3] Wang, Z., Li, S., Chen, X., Li, W., Wu, D., 2025. MVCNet: Multi-view contrastive network for motor imagery classification. Knowledge-Based Systems 328, 114205 <ref:2610.08817#pg4>

[4] Zheng, W.L., Zhu, J.Y., Lu, B.L., 2017. Identifying stable patterns over time for emotion recognition from EEG. IEEE Transactions on Affective Computing 10, 417–429 <ref:2610.08817#pg5>

[5] Wang, Z., Li, S., Wu, D., 2025. Canine EEG helps human: Cross-species and cross-modality epileptic seizure detection via multi-space alignment. National Science Review 12, nwaf086 <ref:2610.08817#pg6>

[6] Ienca, M., Haselager, P., Emanuel, E.J., 2018. Brain leaks and consumer neurotechnology. Nature Biotechnology 36, 805–810 <ref:2610.08817#pg7>

[7] Lin, Z., Zhang, C., Wu, W., Gao, X., 2006. Frequency recognition based on canonical correlation analysis for SSVEP-based BCIs. IEEE Transactions on Biomedical Engineering 53, 2610–2614 <ref:2610.08817#pg8>

[8] Tatum, W.O., Dworetzky, B.A., Schomer, D.L., 2011. Artifact and recording concepts in EEG. Journal of Clinical Neurophysiology <ref:2610.08817#pg9>

[9] Wang, G., Teng, C., Li, K., Zhang, Z., Yan, X., 2015. The removal of EOG artifacts from EEG signals using independent component analysis and multivariate empirical mode decomposition. IEEE Journal of Biomedical and Health Informatics 20, 1301–1308 <ref:2610.08817#pg10>

[10] Shahbakhti, M., Beiramvand, M., Nazari, M., BroniecWójcik, A., Augustyniak, P., Rodrigues, A.S., Wierzchon, M., Marozas, V., 2021. VME-DWT: An efficient algorithm for detection and elimination of eye blink from short segments of single EEG channel. IEEE Transactions on Neural Systems and Rehabilitation Engineering 29, 408–417 <ref:2610.08817#pg11>

[11] Chen, W., Yu, C., Zhao, Z., Zheng, N.

Improvements for AI systems

The DenoFlow architecture enables denoisers to learn the path from a contaminated trial to its clean counterpart rather than a direct mapping between them, allowing for noise suppression that is not constrained by waveform distance alone, as the objective defined on waveform distance alone weigh every sample equally.
The system can perform signal reconstruction via Euler integration of the learned field, as described by the equation Xˆ(s+1) = Xˆ(s) + 1/S fθ(Xˆ(s), X˜′, s/S), which allows for precise forward integration from a contaminated observation.
The classifier supervises the integrated output, ensuring that the denoised signal also has to remain decodable, thereby preserving stimulus-related structure rather than just smoothing away artifacts.
The training process avoids the complexity of adversarial games by optimizing the regression onto the velocity target through a loss function combining flow matching and reconstruction losses: L = Lcls + λflowLflow + λrecLrec.
The system can be trained on real-world physiological noise by injecting known artifacts, such as EMG and EOG signals, under a controlled signal-to-noise target, ensuring the model is robust to practical contamination scenarios.