Review Reports - Personalized Adaptive Gabor Filtering with Three-Stage Semi-Supervised Domain-Adversarial Learning for Cross-Subject SSVEP Decoding

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

A brief summary

The authors proposed a framework for SSVEP decoding that integrates personalized adaptive feature extraction with efficient cross-subject transfer learning. For feature extraction, a set of trainable Gabor filters is used. Their parameters – center frequency and bandwidth – are adapted "per user", i.e., tuned to the individual characteristics of the user's neural response to the stimulating signal. For transfer learning, the authors developed a three-stage lightweight calibration structure, termed TriS-DANN. In the first stage, unsupervised domain alignment is performed. Subsequently, fine-tuning is carried out on a subset of labeled samples from the target domain. With this approach, high cross-subject generalization of the model is achieved using only a small amount of calibration data. Validation of each block of the proposed framework was performed in two experiments: on a public benchmark dataset as well as on the authors' own data collected using a validated methodology from 10 healthy adult male participants aged 21-23 years. In both experiments, the proposed approach achieved better accuracy and information transfer rate. Specifically, accuracy improvements of 4.63% and 3.22% were obtained on the public benchmark dataset and the in-house experimental dataset, respectively, over the conventional fixed filter-bank counterpart. A major achievement of the proposed approach is that with an epoch duration of only 0.4 seconds, TriS-DANN achieved 86.60% accuracy using only 23.07% of the available target-domain training/calibration trials, which demonstrates good practical usability of the BCI. The results of this work have high practical and scientific significance and will be of great interest to the research community.

General concept comments

Article

The manuscript is scientifically sound, and the experimental design is appropriate to test all hypotheses. The manuscript is clear and presented in a well-structured manner. The cited references are relevant, and 16 out of 33 references have been published within the last 5 years. The manuscript contains no self-citations.

Review

The authors should consider the following observation. In the second experiment, the best results were achieved in healthy volunteers with a high SNR. At the same time, the center frequencies of the four Gabor filters were uniformly distributed across the frequency range of 3–50 Hz, and their bandwidths were approximately equal and narrow (6–8 Hz). This raises the question: was adaptive tuning truly necessary in these best-case scenarios, or would a conventional fixed filter bank have performed equally well?

According to the presented data, adaptive tuning was primarily required when the quality of the recorded EEG signal was poor (low SNR). It is known that in healthy volunteers, EEG signal quality (SNR) is determined by a combination of factors: the quality of electrode-to-scalp contact (impedance value), the level of muscle and ocular artifacts, and the effectiveness of power-line noise filtering. Most of these issues are solvable – both in experimental settings and in practical BCI applications, it is possible to achieve an SNR sufficient for reliable decoding. In such cases, the need for adaptation may seem not entirely obvious.

However, in the reviewer's opinion, a fundamentally different situation may be observed in patients with neurodegenerative diseases or those who have suffered a stroke. It is precisely for this population that the proposed user-specific adaptation may prove critically important. If the authors' findings are consistent with this view, the Discussion section could be expanded accordingly. Additionally, the authors should specify what type of electrodes were used in the study – dry or wet (gel-based).

Specific comments

The abbreviation TriS-DANN used in the title is not widely recognized in the BCI community. In contrast to the well-established abbreviation SSVEP, which readers readily understand, TriS-DANN (standing for "three-stage semi-supervised domain-adversarial neural network") is not a standard term. This lack of clarity makes the title less transparent and reduces the work's searchability. The reviewer strongly recommends revising the title to either fully spell out the model's name or use an abbreviation that is clearly defined and widely accepted.
In line 204 (section 2.2.2), the specified frequency range (8–13) Hz seems incorrect. Based on the article, four harmonics (n = 4) were taken into account within the filtered EEG range of (3–50) Hz. Otherwise, the sentence is ambiguous and should be clarified.
The authors should revise their variable notation. Throughout the article, different variables are denoted by the same letter:

In formula (1), trepresents time (where t∈[−0.5(K−1),0.5(K−1)]), and s denotes the filter index corresponding to the learned center frequency fs (where s=1,2,…,S).
However, in formula (5), tj is used to denote the j-th target band (key frequency band). Moreover, the same formula (5) in Algorithm 1 uses ti instead of tj.
In Algorithm 2, the indices sand t are used as abbreviations for "Source" and "Target" and appear as domain subscripts, e.g., Ds, Dt, among others.

To improve readability and avoid confusion, it is advisable to use unique notation for different variables, constants, abbreviations, and other elements when presenting mathematical formalizations.

Figure 6 seems to contain a typo. Specifically, the label "C-AFE-TCNN" is used, but based on the context, it should likely read "G-AFB-tCNN". Please verify and correct accordingly.
The authors should add a reference to Figure 3 in the text, as it is currently missing.

Author Response

Responses to Reviewer 1

Comments: The authors proposed a framework for SSVEP decoding that inte-

grates personalized adaptive feature extraction with efficient cross-subject transfer

learning. For feature extraction, a set of trainable Gabor filters is used. Their parameters– center frequency and bandwidth–are adapted “per user”, i.e., tuned to the individ-

ual characteristics of the user’s neural response to the stimulating signal. For trans- fer learning, the authors developed a three-stage lightweight calibration structure, termed TriS-DANN. In the first stage, unsupervised domain alignment is performed.

Subsequently, fine-tuning is carried out on a subset of labeled samples from the tar- get domain. With this approach, high cross-subject generalization of the model is achieved using only a small amount of calibration data. Validation of each block of the proposed framework was performed in two experiments: on a public benchmark dataset as well as on the authors’ own data collected using a validated methodology from 10 healthy adult male participants aged 21-23 years. In both experiments, the proposed approach achieved better accuracy and information transfer rate. Specifi- cally, accuracy improvements of 4.63% and 3.22% were obtained on the public bench- mark dataset and the in-house experimental dataset, respectively, over the conven- tional fixed filter-bank counterpart. A major achievement of the proposed approach is that with an epoch duration of only 0.4 seconds, TriS-DANN achieved 86.60% ac- curacy using only 23.07% of the available target-domain training/calibration trials, which demonstrates good practical usability of the BCI. The results of this work have high practical and scientific significance and will be of great interest to the research community.

The manuscript is scientifically sound, and the experimental design is appro- priate to test all hypotheses. The manuscript is clear and presented in a well- structured manner. The cited references are relevant, and 16 out of 33 references have been published within the last 5 years. The manuscript contains no self- citations.

Reply: We thank the reviewer for the constructive evaluation and insightful com- ments, which have significantly strengthened the rigor of our manuscript.

To facilitate the review process, we present our point-by-point responses in a color-coded format. Our responses to the reviewers’ comments are shown in boxes with a light blue background , while the corresponding textual revisions from the manuscript are shown in boxes with a light green background . In addition, all modifications made in the revised manuscript are marked with red underlines.

In response to the reviewer’s comments, we have revised the manuscript in sev- eral aspects. For Comment 1.1, we clarified the practical value and boundary con- ditions of the proposed G-AFB, emphasizing that adaptive filtering is particularly useful under unstable EEG quality or pronounced inter-subject variability, and we also specified the use of wet electrodes with conductive paste in the in-house ex- periment. For Comment 1.2, we revised the title by removing the non-standard abbreviation “TriS-DANN” and spelling out the main methodological components more clearly. For Comment 1.3, we clarified the relationship among the fundamen- tal stimulus-frequency range, the EEG analysis range, and the harmonic order used for G-AFB initialization. For Comment 1.4, we revised the variable notation in the G-AFB and TriS-DANN formulations to improve consistency and avoid ambiguity. For Comment 1.5, we corrected the typographical error in Figure 6 by replacing “C- AFE-tCNN” with “G-AFB-tCNN”. Finally, for Comment 1.6, we added the missing in-text reference to Figure 3 and improved the connection between the figure and the corresponding methodological description.

Comment:

1.1 The authors should consider the following observation. In the second experiment, the best results were achieved in healthy volunteers with a high SNR. At the same time, the center frequencies of the four Gabor filters were uniformly distributed across the frequency range of 3–50 Hz, and their bandwidths were approximately equal and narrow (6–8 Hz). This raises the question: was adaptive tuning truly necessary in these best-case scenarios, or would a conventional fixed filter bank have performed equally well?

According to the presented data, adaptive tuning was primarily required when the quality of the recorded EEG signal was poor (low SNR). It is known that in healthy volunteers, EEG signal quality (SNR) is determined by a combination offactors: the quality of electrode-to- scalp contact (impedance value), the level of muscle and ocular artifacts, and the effectiveness of power-line noise filtering. Most of these issues are solvable – both in experimental set- tings and in practical BCI applications, it is possible to achieve an SNR sufficient for reliable decoding. In such cases, the need for adaptation may seem not entirely obvious.

However, in the reviewer’s opinion, a fundamentally different situation may be observed in patients with neurodegenerative diseases or those who have suffered a stroke. It is precisely for this population that the proposed user-specific adaptation may prove critically important. If the authors’findings are consistent with this view, the Discussion section could be expanded accordingly. Additionally, the authors should specify what type of electrodes were used in the study–dry or wet (gel-based).

Reply:

We thank the reviewer for this insightful comment. We agree that, under ideal record- ing conditions in healthy participants with high SNR, a conventional fixed filter bank may already provide strong decoding performance, and the additional benefit of adaptive tuning may be relatively modest. Our intention is therefore not to claim that adaptive filtering is always necessary under best-case conditions, but to show that a learnable filter bank provides a user-specific mechanism for adapting to inter- subject variability in SSVEP responses.

This interpretation is consistent with our results on the in-house dataset. In Sec- tion 3.2.3, G-AFB-tCNN improved the 0.4 s decoding accuracy from 88.63% with

FB-tCNN to 91.85%, indicating a measurable improvement over the fixed filter-bank counterpart. More importantly, the visualization analysis in Section 3.2.4 and Figure 10 shows that the learned filters behaved differently across subjects with different signal quality. For S01, who had high SNR, the learned center frequencies were close to a regular harmonic structure, and the bandwidths were relatively narrow and con- sistent. By contrast, for S03, who had lower signal quality, G-AFB learned a wider passband at the fundamental frequency, suggesting an adaptive compensation strat- egy for unstable or low-SNR responses.

Following the reviewer’s suggestion, we have revised the Discussion to clarify this point. We now state that the benefit of G-AFB may be limited when SSVEP re- sponses are stable and have high SNR, because a fixed filter bank may already be sufficient in such cases. However, user-specific adaptive filtering may be more valu- able when EEG signal quality is unstable, calibration data are limited, or inter-subject variability is more pronounced. We have also added a discussion that this property may be particularly relevant to populations such as patients with neurodegenerative diseases or stroke, although further validation in such populations is required.

In addition, we have revised the in-house dataset description to clarify that wet, conductive-paste electrodes were used with an iRecorder W32 wireless EEG ampli- fier.

Changes:

In Section 3.1.2. In-House Dataset:

Visual stimuli were programmed using PsychoPy (version 2025.1.1) in Python 3.9 and presented on a 27-inch LCD monitor with a 165 Hz refresh rate to ensure stable stimulation frequencies. EEG signals were recorded using wet electrodes with conductive paste and an iRecorder W32 wireless EEG/ERP acquisition system from

Shanghai Niantong Intelligent Technol-ogy Co., Ltd. Eight channels located in or

near the occipital region were recorded: Oz, O1, O2, PO3, PO4, P7, Pz, and P8, as

shown in Figure 3.

In Section 4.1 Mechanism of Individualized Adaptation in G-AFB:

decoding performance by enabling individualized adaptation. This advantage

stems from its ability to respond to physiological differences among users. In our

experiments, such differences were reflected by SNR variability; however, SNR is

only one aspect of inter-subject differences. From a neurophysiological perspective,

factors such as skull thickness, cortical structure, and neuronal organization

may shift the energy distribution, phase characteristics, and optimal response

frequency bands of SSVEP responses. Traditional fixed filter banks are based

on a one-size-fits-all assumption and therefore cannot capture these subtle but

important individual characteristics, limiting decoding performance. In contrast, the

data-driven G-AFB layer can dynamically adjust its center frequency and bandwidth

through end-to-end learning, thereby matching each user’s neural response pattern

and improving-decoding-performance.

This interpretation is further supported by the visualization results for subjects with different signal quality. In high-SNR participants such as S01, the learned

filters approached a regular harmonic structure, indicating that a fixed filter bank

may already be effective under favorable recording conditions. By contrast, for

the lower-SNR subject S03, G-AFB learned a wider passband at the fundamental

frequency, suggesting adaptive compensation for weaker or less stable SSVEP

responses. Thus, G-AFB may be particularly useful when EEG quality is

unstable or individual variability is pronounced. This property may be relevant

to populations with greater neurophysiological heterogeneity, such as patients

with neurodegenerative diseases or stroke, although further clinical validation is

required.

Comment:

1.2 The abbreviation TriS-DANN used in the title is not widely recognized in the BCI com- munity. In contrast to the well-established abbreviation SSVEP, which readers readily un- derstand, TriS-DANN (standing for “three-stage semi-supervised domain-adversarial neural network”) is not a standard term. This lack of clarity makes the title less transparent and reduces the work’s searchability. The reviewer strongly recommends revising the title to ei- ther fully spell out the model’s name or use an abbreviation that is clearly defined and widely

accepted.

Reply:

We thank the reviewer for this helpful suggestion. We agree that “TriS-DANN” is a model-specific abbreviation rather than a widely recognized term in the BCI community. Therefore, using this abbreviation directly in the title may reduce the transparency and searchability of the manuscript, especially for readers who are not familiar with the proposed framework.

To address this issue, we have revised the title by removing the abbreviation “TriS- DANN” and replacing it with a descriptive phrase that explicitly states the main methodological components of the study. The revised title is: Personalized Adaptive Gabor Filtering with Three-Stage Semi-Supervised Domain-Adversarial Learning for Cross-Subject SSVEP Decoding.

This revised title preserves the two central contributions of the manuscript, namely personalized adaptive Gabor filtering and three-stage semi-supervised domain-adversarial learning, while avoiding an undefined non-standard abbrevia- tion in the title. In addition, TriS-DANN is now defined in full at its first occurrence in the Abstract and Methods section, and the abbreviation is used consistently there- after.

Comment:

1.3 In line 204 (Section 2.2.2), the specified frequency range (8–13 Hz) seems incorrect. Based on the article, four harmonics (n=4) were taken into account within the filtered EEG range of 3–50 Hz. Otherwise, the sentence is ambiguous and should be clarified.

Reply:

We thank the reviewer for pointing out this ambiguity. We agree that the orig- inal wording may have caused confusion because the range 8–13 Hz refers only to the fundamental stimulus-frequency range, rather than to the complete EEG analysis range or the full frequency range covered by the Gabor filters.

We have revised Section 2.2.2 to clarify this point. In the revised manuscript, we explicitly state that the EEG analysis range was 3–50 Hz and that the first four

harmonics (n=4) of the stimulus frequencies were considered. The initial center fre- quency of each Gabor subband was determined from the corresponding harmonic range, where fmin and fmax denote the minimum and maximum stimulus frequencies, respectively. This revision clarifies the relationship among the fundamental stimulus- frequency range, the EEG analysis range, and the harmonic order used in the G-AFB initialization.

Changes:

In Section 2.2.2. Loss Function Regularization:

To constrain the learned-——Gabor-kernel parameters within—physiologically meaningful ranges and accelerate convergence, we use a guidance mechanism

that combines parameter initialization with prior-based regularization. Instead of

random initialization, the initial parameters of each subband filter were determined

according to the fundamental stimulus-frequency range of 8—13 Hz and its

harmonic components.

Specifically,-the E——EGanalysis range was 3–50 Hz, and the first four harmonics (n = 4) of the stimulus frequencies were considered. In this study, the number

of Gabor subbands was set to S = 4, corresponding to the first four harmonic

ranges of the stimulus frequencies. For the n-th harmonic band, the initial center

frequency was set to the theoretical center of the corresponding harmonic range,

(fmin × n + fmax × n)/2, where fmin and fmax denote the minimum and maximum

stimulus frequencies, respectively. Therefore, each learnable Gabor kernel was

initialized around one target-related harmonic band, and its center frequency

and bandwidth were further optimized during training through end-to-end

Comment:

1.4 The authors should revise their variable notation. Throughout the article, different vari- ables are denoted by the same letter.

In Formula (1), t represents time, where t∈ [−0.5(K − 1), 0.5(K − 1)], and s denotes the filter index corresponding to the learned center frequency fs, where s = 1, 2, . . . , S.

However,in Formula (5), tj is used to denote the j-th target band, namely the key fre- quency band. Moreover, in Algorithm 1, the same Formula (5) uses ti instead of tj.
In Algorithm 2, the indices s andt are used as abbreviations for “Source” and “Target” and appear as domain subscripts,for example, Ds and Dt.

To improve readability and avoid confusion, it is advisable to use unique notation for dif- ferent variables, constants, abbreviations, and other elements when presenting mathematical formalizations.

Reply:

We thank the reviewer for pointing out this important notation issue. We agree that the original manuscript used several overlapping symbols, which could reduce the readability of the mathematical formulation.

To address this issue while avoiding unnecessary changes to standard local nota- tion in the Gabor-kernel definition, we revised the notation according to the princi- ples summarized in the table below. Specifically, we retained t as the local time index and s as the Gabor filter/subband index in Formula (1). We then revised the ambigu- ous cross-section notation: the target-related frequency bands are now uniformly de- noted by bj, and the source/target-domain datasets and samples are denoted by Dsrc, D tar, xsrc, and xtar, respectively.

Original

symbol

Original meaning

Issue

Revision

Local time index in

The notation itself is

Retained as t; explicitly

the Gabor kernel

standard, but it may be confused with the target- domain subscript if not

defined before Formula (1) as the local time index.

explicitly defined.

s	Gabor fil ter/subband index	The notation itself is standard, but it may be confused with the source- domain subscript in Ds .	Retained as s; explicitly defined before Formula (1) as the filter/subband index, s = 1, 2, . . . , S.
fs	Center frequency of the s-th filter	May be confused with a source-domain subscript if Ds is also used.	Retained as fs because s is the filter index; Ds is no longer used for the source domain.
σs	Bandwidth of the s- th filter	Same as above.	Retained as σs .
tj, ti	The j-th or i-th key frequency band	Conflicts with the time vari- able t; Formula (5) and Al- gorithm 1 used inconsistent notation.	Unified as bj.
Ds, Dt	Source/target- domain datasets	Lowercase s and t may be confused with the filter in- dex s and the local time variable t.	Revised as Dsrc and D tar.
Xs, Xt	Source/target sam-	May also be confused with	Revised as Xsrc and X tar,
Xs, Xt	ples or features	the indices s and t.	or sample-wise notation such as xisrc and xjtar.
Ys	Feature map of the	May appear similar to	Retained as Ys .
Ys	s-th subband	source-domain notation,

but it is internally consis-

tent with gs, fs, and σs in

G-AFB.

These revisions have been applied consistently in Section 2.2, Section 2.3. The two revised algorithms are also shown in the changes below for clarity.

Changes:

The revised Algorithm 1 and Algorithm 2 are shown below.

Algorithm 1: Forward Propagation and Loss Calculation of the G-AFB Layer

Require: EEG batch X ∈ RN×C×T, true labels Y true, downstream model DL Model, filter number S, target bands {bj}, weights λ, α, β

Require: Learnable parameters {fs }, {σs}, θDL

Ensure: Total loss L total

1: Initialize {fs } and {σs} with heuristic rules

2: for each filter s = 1, . . . , S do

3: Generate Gabor kernel gs(t) with (fs, σs)

4: Apply depthwise convolution of X with gs(t), obtain Ys

5: end for

6: Concatenate {Ys } along filter dimension → YG-AFB

7: Ypred ← DL Model(YG-AFB)

8: Compute classification loss:

Lcls ← CrossEntropy (Ypred, Y true )

9: Compute prior losses:

L prior = λ (Lfreq+ αLbw+ βL diff ) .

10: L total ← Lcls + Lprior

11: return L total

Algorithm 2: Three-Stage Semi-Supervised Domain-Adversarial Training Procedure

Require: Source-domain data Dsrc = {(xisrc, yisrc)}

Require: Target-domain labeled data (few-shot) Dtarl = {(xktar, yktar)}

Require: Feature extractor F(·; θF ), label classifier Cy(·; θy ), Cd(·; θd)

Ensure: Adapted feature extractor θF, label classifier θy, domain classifier θd

Ensure: Final adapted model M = {F, Cy} Step 1: Source-Domain Pre-Training

1: Initialize θF, θy, and θd

2: for minibatch (xisrc, yisrc) ~ Dsrc do

3: fisrc ← F(xisrc; θF ); ypred ← Cy (fisrc; θy )

4: Compute classification loss:

L label = CrossEntropy (ypred, yisrc)

5: Update θF and θy by minimizing L label

6: end for

Step 2: Unsupervised Domain Adaptation

7: for minibatch (xisrc, yisrc) ~ Dsrc and minibatch xjtar ~ Dtaru do

8: fisrc ← F(xisrc; θF ); far ← F(xjtar; θF )

9: Pass features through Gradient Reversal Layer (GRL)

10: dpsrercd ← Cd(fisrc; θd); dpretard← Cd(far; θd)

11: Compute domain loss:

L domain = CrossEntropy (dpsrercd, source) + CrossEntropy (dpretard, target)

12: Update θd to minimize L domain (domain discrimination)

13: Update θF via GRL to maximize L domain (domain confusion)

14: end for

Note: θy is frozen, and only θF and θd are updated.

Step 3: Fine-Tuning with Few Labeled Samples

15: for minibatch (xktar, yktar) ~ Dtarl do

16: fktar ← F(xktar; θF )

17: ypretard← Cy (fktar; θy )

18: Compute fine-tuning loss:

L fine = CrossEntropy (ypretard, yktar)

19: Update θF and θy by minimizing L fine

20: end for

21: return θF, θy, θd

Comment:

1.5 Figure 6 seems to contain a typo. Specifically, the label “C-AFE-TCNN” is used, but based on the context, it should likely read “G-AFB-tCNN”. Please verify and correct accord- ingly.

Reply:

We thank the reviewer for carefully identifying this typographical error. We have verified that the label “C-AFE-tCNN” in Figure 6 was incorrect. The intended label is “G-AFB-tCNN”, which is consistent with the model name used throughout the manuscript.

Changes:

Figure 6: Average classification accuracy on the in-house dataset under different time-window lengths. Error bars indicate standard deviation.

Comment:

1.6 The authors should add a reference to Figure 3 in the text, as it is currently missing.

Reply:

We thank the reviewer for pointing out this omission. We have added an explicit in-text reference to Figure 3 in Section 3.1.2, where the in-house EEG acquisition setup and selected channels are described. The revised text now refers to Figure 3 when introducing the eight-channel occipital and parieto-occipital montage, thereby linking the figure more clearly with the corresponding methodological description.

Changes:

In Section 3.1.2. In-House Dataset:

Shanghai Niantong Intelligent Technology Co., Ltd. Eight channels located in or near

the occipital region were recorded: Oz, O1, O2, PO3, PO4, P7, Pz, and P8, as shown

in Figure 3.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript proposes two main components for SSVEP decoding: a Gabor adaptive filter bank (G-AFB), which aims to replace conventional fixed filter banks with learnable filtering parameters, and a three-stage semi-supervised domain-adversarial network (TriS-DANN), which aims to reduce calibration requirements in cross-subject SSVEP decoding. The abstract reports that G-AFB-tCNN achieved 89.13% accuracy on the public dataset, 91.85% accuracy on the in-house dataset with a 0.4 s window, and that TriS-DANN achieved 86.60% accuracy using limited target-domain data. These results are potentially interesting. However, I have several major concerns regarding the experimental design, data partitioning, reproducibility, and clarity of the method description. The reported practical implications should therefore be reconsidered.

Major Comments

The in-house experimental design does not adequately support the claim of fast short-window SSVEP-BCI decoding.

A major concern is that the in-house experiment appears to use a block design, in which each block corresponds to one fixed stimulus frequency or idle state. Moreover, each trial contains a long stimulation period of 12.14 s. This design is substantially different from a standard trial-by-trial SSVEP-BCI selection paradigm, where the target changes from trial to trial and decoding is performed shortly after stimulus onset.

Although the authors report high accuracy using short windows such as 0.4 s or 1.0 s, these windows appear to be randomly extracted from the long 12 s stimulation epoch. Therefore, a 0.4 s segment may correspond to a period after the participant has already viewed the same stimulus for several seconds. Such a segment is likely influenced by the preceding long-term steady visual stimulation and does not represent the first 0.4 s of SSVEP response after stimulus onset. Consequently, the in-house dataset cannot be used as strong evidence that the proposed method enables fast SSVEP-BCI decoding in a practical short-window setting.

To support claims about rapid SSVEP decoding, the authors should report results using onset-locked windows, for example, only the first 0.4 s after visual latency correction. Preferably, this analysis should be performed on the public benchmark dataset using the earliest post-stimulus segment, rather than selecting arbitrary short segments from a long stimulation period. The current in-house results should be interpreted more cautiously as classification of short segments extracted from stabilized long-duration SSVEP responses, not as evidence of rapid BCI operation.

The random-start sliding-window strategy may cause data leakage or performance overestimation.

The manuscript uses a random-start sliding-window strategy to improve temporal robustness and generalization. However, this strategy can also introduce serious data leakage if it is not handled carefully. If multiple overlapping or highly similar windows are generated from the same original trial before the training/test split, windows derived from the same trial may appear in both the training and test sets. This would artificially inflate the reported accuracy.

Moreover, the authors should consider the requirements of real BCI applications. In a practical SSVEP-BCI system, decoding is usually expected to occur shortly after stimulus onset. Therefore, randomly selected short segments from a long stimulation epoch should not be treated as equivalent to the first 0.4 s or 1.0 s after stimulus onset. To demonstrate fast SSVEP decoding, the authors should perform an onset-locked analysis using only the early post-stimulus segment of each trial and ensure that all data splitting is performed at the trial or block level before any sliding-window augmentation is applied.

The manuscript proposes an algorithm but does not provide sufficient code availability information.

The proposed framework includes several implementation-sensitive components, including the G-AFB parameterization, prior regularization, three-stage TriS-DANN training, and model selection. Although the authors provide a general description, these details may not be sufficient for independent reproduction.

The authors should release the source code and training scripts, or at least provide a clear code availability statement explaining why the code cannot be made public. For an algorithm-oriented deep learning paper, code availability is important for reproducibility and for promoting the method’s broader use.

The figures lack sufficiently detailed legends and explanations.

Most figures are not self-contained and lack detailed legends. This reduces readability and makes it difficult to understand the experimental protocol and data flow. This is especially problematic for Figure 1. In Step 2, the figure shows N-1 subjects as the source domain and subject N as the target domain, but it does not clearly distinguish which part of the target data is used for unsupervised adaptation, which part is used for fine-tuning, which part is used for validation, and which part is reserved for testing. If the figure is read without the main text, it may appear that the target subject’s data are used without proper separation, raising concerns about data leakage.

The authors should revise all figure legends to make the figures understandable without relying heavily on the main text. In particular, Figure 1 should explicitly indicate the source-domain data, target-domain unlabeled adaptation data, labeled calibration data, validation data, and final test data.

Minor Comments

Line 177: “a set of shared learnable Gabor kernels” is unclear.
Please clarify what “a set” refers to. How many Gabor kernels are included? What does “shared” mean here? Are the kernels shared across channels, trials, subjects, or models?
Line 181: The selection or initialization of fs should be explained.
The manuscript states that fs is the center frequency of the Gabor kernel, but it is not sufficiently clear how fs is initialized, constrained, or updated during training.
Line 205: The definitions of fmin, fmax, and n are incomplete.
Please specify whether fmin and fmax correspond to the minimum and maximum stimulus frequencies. Also clarify what n represents, whether it is the harmonic order, and what values of n are used.
Line 221: The predefined bandwidth range [2, 15] Hz needs justification.
The manuscript describes [2, 15] Hz as a physiologically plausible range, but the basis for this choice is unclear. Please explain whether this range is based on prior literature, empirical tuning, or physiological assumptions.
The description of short-window analysis should be revised.
The authors should avoid implying that randomly selected 0.4 s windows from a long stimulation epoch are equivalent to the first 0.4 s after stimulus onset. These two analyses have different implications for practical BCI systems.

Author Response

Responses to Reviewer 2

Comments: The manuscript proposes two main components for SSVEP decod-

ing: a Gabor adaptive filter bank (G-AFB), which aims to replace conventional fixed

filter banks with learnable filtering parameters, and a three-stage semi-supervised

domain-adversarial network (TriS-DANN), which aims to reduce calibration require-

mentsin cross-subjectSSVEP decoding. The abstract reports that G-AFB-tCNN achieved 89.13% accuracy on the public dataset, 91.85% accuracy on the in-house dataset with

a 0.4 s window, and that TriS-DANN achieved 86.60% accuracy using limited target- domain data. These results are potentially interesting. However, I have several major concerns regarding the experimental design, data partitioning, reproducibility, and clarity of the method description. The reported practical implications should there- fore be reconsidered.

Reply: We sincerely thank the reviewer for the insightful comments, which have been instrumental in enhancing the overall quality of our manuscript.

In response to the reviewer’s comments, we have revised the manuscript in sev- eral aspects. For Comment 2.1, we clarified the overall interpretation of the proposed G-AFB and TriS-DANN framework, with particular attention to experimental design, data partitioning, reproducibility, and the practical implications of the reported short- window results. For Comment 2.2, we clarified that train, validation, and test par- titions were performed at the original trial/epoch level before random-start sliding- window generation, thereby preventing data leakage between subsets. For Comment 2.3, we provide the source code and related training scripts through the following GitHub repository: https://github.com/Xinyu928-hub/sensors-4331333-. We also

added a clearer Data and Code Availability Statement to explain how the public benchmark dataset, source code, and in-house EEG data can be accessed. For Com- ment 2.4, we revised Figure 1 and its legend to explicitly distinguish source-domain data, target-domain unlabeled adaptation data, labeled calibration data, validation data, and final test data, and we expanded other figure legends to improve self- contained readability. For Comment 2.5, we clarified that the G-AFB layer uses S = 4 shared learnable Gabor kernels and explained that “shared” means the same kernel parameters are applied across EEG channels and trials within the same model. For Comment 2.6, we added details on the initialization, constraint, and backpropagation- based update of the learnable center frequency fs . For Comment 2.7, we explicitly defined f min, fmax, and the harmonic order n, and clarified that the first four harmon- ics were considered in this study. For Comment 2.8, we explained the rationale for the predefined bandwidth range of [2,15] Hz, noting that it was selected based on physiological plausibility and empirical tuning. Finally, for Comment 2.9, we further revised the Methods and Discussion to distinguish random short-segment analysis from onset-locked early post-stimulus analysis and stated this limitation explicitly.

Comment:

2.1 1. The in-house experimental design does not adequately support the claim of fast short- window SSVEP-BCI decoding.

A major concern is that the in-house experiment appears to use a block design, in whicheach block corresponds to one fixed stimulus frequency or idle state. Moreover, each trial contains a long stimulation period of 12.14 s. This design is substantially different from a standard trial-by-trial SSVEP-BCI selection paradigm, where the target changes from trial to trial and decoding is performed shortly after stimulus onset.
Although the authors report high accuracy using short windows such as0.4 s or 1.0 s, these windows appear to be randomly extracted from the long 12 s stimulation epoch. Therefore, a 0.4 s segment may correspond to a period after the participant has already viewed the same stimulus for several seconds. Such a segment is likely influenced by the preceding long-term steady visual stimulation and does not represent the first 0.4 s of SSVEP response after stimulus onset. Consequently, the in-house dataset cannot be used as strong evidence that the proposed method enables fast SSVEP-BCI decoding in a practical short-window setting.
To support claims about rapid SSVEP decoding, the authors should report results usingonset-lockedwindows, for example, only the first 0.4 s after visual latency correction. Preferably, this analysis should be performed on the public benchmark dataset using the earliest post-stimulus segment, rather than selecting arbitrary short segments from a long stimulation period. The current in-house results should be interpreted more cautiously as classification of short segments extracted from stabilized long-duration SSVEP responses, not as evidence of rapid BCI operation.

Reply:

We thank the reviewer for raising this important concern. We agree that the in- house experiment should not be interpreted as a standard trial-by-trial online SSVEP- BCI selection paradigm. In our in-house dataset, each task block corresponded to one stimulus frequency or the idle state, and each trial contained a relatively long stimulation period. Therefore, a randomly extracted 0.4 s or 1.0 s segment may come from a stabilized SSVEP response after the participant had already viewed the same stimulus for several seconds. Such a segment is not equivalent to the first 0.4 s or 1.0 s after stimulus onset.

Accordingly, we have revised the manuscript to avoid overinterpreting the in- house short-window results as direct evidence of rapid online BCI operation. We now describe these results as classification performance using short segments extracted from the stimulation period, or from stabilized long-duration SSVEP responses. This interpretation is more consistent with the experimental design and avoids implying that the in-house dataset directly supports onset-locked fast BCI selection.

We agree that an onset-locked analysis using the earliest post-stimulus segment would be valuable for evaluating true rapid SSVEP decoding. However, the cur- rent in-house experiment was designed as a block-based long-stimulation paradigm rather than a trial-by-trial online selection paradigm. Therefore, it is not appropriate to use this dataset as strong evidence for onset-locked rapid BCI operation. Instead of adding an unsupported claim, we have revised the Abstract, Methods, Results, and Discussion to clarify the window-extraction strategy and to state this limitation explicitly. We also note in the revised Discussion that future work should validate the proposed framework in a strict trial-by-trial online paradigm using onset-locked windows after visual-latency correction.

Changes:

In the Abstract:

In transfer learning, TriS-DA——NNreached-86.60-—%accuracy using-0.4 s segments extracted from the stimulation period and only 23.07% of the available target-domain

training/calibration trials, demonstrating higher efficiency and stability than

conventional fine-tuning.

In Section 3.1.2. In-House Dataset:

delay, the EEG segment from 0.14 to 12.14 s after stimulus onset was extracted as the

analysis epoch, corresponding to a 12.0 s period of sustained visual stimulation.

In Section 3.2.1. Experimental Setup:

samples from each 12.0 s analysis epoch. For each predefined window

length, the starting point was randomly selected within the analysis epoch,

and the corresponding EEG segment was extracted as one sample (Figure 5).

Importantly, these randomly selected short windows should not be interpreted as

onset-locked early post-stimulus windows. Because the in-house experiment used a

long-stimulation block design, a randomly selected 0.4 s or 1.0 s segment may come

from a stabilized response period rather than from the first 0.4 s or 1.0 s after stimulus

onset.

In Section 3.2.3. Performance Evaluation on the In-House Dataset with an Idle State:

As shown in Figure 6, the accuracy of all models increased as the time window became longer, which is consistent with expectations. G-AFB-tCNN achieved the

best performance under all tested conditions. In the most challenging 0.4 s segment

setting, G-AFB-tCNN reached an average accuracy of 91.85%, clearly outperforming

the other models and demonstrating strong classification performance for short

segments extracted from stabilized SSVEP responses. Its ITR also peaked at 0.4 s

(Figure 7), indicating a favorable balance between segment length and classification

accuracy in the present offline analysis.

In Section 4.3 Limitations and Future Work:

It should be noted that the in-house short-window results should not be interpreted as direct evidence of onset-locked rapid online BCI operation. The

in-house experiment used a block-based long-stimulation design, and the 0.4 s and

1.0 s samples were extracted from the 12.0 s analysis epoch after visual-latency

correction. Therefore, these samples may correspond to stabilized SSVEP responses

rather than the earliest post-stimulus response. The present results demonstrate

the ability of the proposed method to classify short segments extracted from

long-duration SSVEP responses, but further validation using a strict trial-by-trial

online paradigm and onset-locked windows is required to evaluate true rapid

SSVEP-BCI decoding.

Comment:

2.2 The random-start sliding-window strategy may cause data leakage or performance over- estimation.

• The manuscript uses a random-start sliding-window strategy to improve temporal ro- bustness and generalization. However, this strategy can also introduce serious data leakage if it is not handled carefully. If multiple overlapping or highly similar windows are generated from the same original trial before the training/test split, windows derived from the same trial may appear in both the training and test sets. This would artificially inflate the reported accuracy.

• Moreover, the authors should consider the requirements of real BCI applications. In a practical SSVEP-BCI system, decoding is usually expected to occur shortly after stim- ulus onset. Therefore, randomly selected short segments from a long stimulation epoch should not be treated as equivalent to the first 0.4 s or 1.0 s after stimulus onset. To demonstrate fast SSVEP decoding, the authors should perform an onset-locked analysis using only the early post-stimulus segment of each trial and ensure that all data split- ting is performed at the trial or block level before any sliding-window augmentation is applied.

Reply:

We thank the reviewer for raising this important concern. We agree that the random-start sliding-window strategy may lead to data leakage and performance overestimation if the window samples are generated before data partitioning. In par- ticular, if multiple overlapping or highly similar windows derived from the same original trial are distributed across the training and test sets, the test accuracy would no longer reflect an independent evaluation.

To avoid this issue, the data partitioning in our experiments was performed at the original trial/epoch level before applying the sliding-window strategy. The random-

start sliding-window procedure was then conducted independently within each sub- set. Therefore, all short-window samples derived from the same original trial/epoch were assigned to the same subset, and no window generated from a test trial/epoch was used for training, validation, model selection, or hyperparameter tuning. We have revised the manuscript to explicitly clarify this procedure. The source code and related training scripts are available at https://github.com/Xinyu928-hub/sensors- 4331333-.

We also agree with the reviewer that randomly selected short segments from a long stimulation epoch should not be regarded as equivalent to the earliest 0.4 s or 1.0 s response after stimulus onset. Accordingly, we have revised the manuscript to avoid interpreting the in-house short-window results as direct evidence of onset- locked rapid online BCI operation. The revised text now states that these results demonstrate the classification of short segments extracted from long-duration SSVEP responses. We further added this issue as a limitation and clarified that future work should validate the proposed method using a strict trial-by-trial online paradigm with onset-locked post-stimulus windows.

Changes:

In Section 3.2.1. Experimental Setup:

All train/validation/test-partition——swere performed at-the original trial/epoch level before applying the random-start sliding-window procedure. After data

partitioning, short-window samples were generated independently within each

subset. Consequently, all windows derived from the same original trial/epoch

remained in the same subset, and no window from the test set was used for training,

validation, model selection, or hyperparameter tuning.

long-stimulation block design, a randomly selected 0.4 s or 1.0 s segment may come
from a stabilized response period rather than from the first 0.4 s or 1.0 s after stimulus
onset. In Section 4.3. Limitations and Future Work: It should-be noted th--- -atthe in-house short-window results should not-be interpreted as direct evidence of onset-locked rapid online BCI operation. The
in-house experiment used a block-based long-stimulation design, and the 0.4 s and
1.0 s samples were extracted from the 12.0 s analysis epoch after visual-latency
correction. Therefore, these samples may correspond to stabilized SSVEP responses
rather than	the earliest post-stimulus	response.	The present results demonstrate
the ability	of the proposed method	to classify	short segments extracted from
long-duration SSVEP responses, but further validation using a strict trial-by-trial
online paradigm and onset-locked windows is required to evaluate true rapid
SSVEP-BCI decoding.

Comment:

2.3 The manuscript proposes an algorithm but does not provide sufficient code availability information.

Theproposed frameworkincludes several implementation-sensitive components, in- cluding the G-AFB parameterization, prior regularization, three-stage TriS-DANN training, and model selection. Although the authors provide a general description, these details may not be sufficient for independent reproduction.
Theauthorsshould release the source code and training scripts, or at least provide a clear code availability statement explaining why the code cannot be made public. For an algorithm-oriented deep learning paper, code availability is important for reproducibil- ity and for promoting the method’s broader use.

Reply:

We thank the reviewer for this constructive suggestion. We agree that code avail- ability is important for the reproducibility and broader use of an algorithm-oriented deep learning method. To address this concern, we have revised the manuscript from two aspects.

First, during the review process, we provide access to the source code of the pro- posed framework, including the implementation of G-AFB-tCNN and TriS-DANN, as well as the related training scripts and configuration files, through the following GitHub repository: https://github.com/Xinyu928-hub/sensors-4331333-. Because the code is associated with an ongoing research project, subsequent access and reuse will be handled by the corresponding author. After publication, readers who need the source code for academic or reproducibility purposes may contact the corresponding author upon reasonable request. The in-house dataset is not publicly released be- cause it was collected as part of the same project and contains human-subject EEG recordings. The data may be provided upon reasonable request, subject to institu- tional approval, project-related restrictions, and participant privacy protection re- quirements. The revised manuscript now includes clearer Code Availability and Data Availability statements indicating how the code and data can be accessed.

Second, we have further expanded the implementation details in the manuscript. In particular, we have added more detailed descriptions of the G-AFB parameteri- zation, the prior regularization term, the three-stage TriS-DANN training procedure, model selection, and the main hyperparameter settings. These revisions are intended to make the proposed method easier to reproduce independently, even when readers follow the manuscript description directly.

Changes:

In Section 2.2.1 Gabor Kernels and Adaptive Mechanism, we further clarified the learnable G-AFB parameters and their constraints:

This kernel function is-determined by two learnable parameters:--the center frequency fs ∈ [3, 50] Hz, which specifies the spectral center of the subband of

interest, and the subband bandwidth σs ∈ [0.5, 25] Hz, which determines the

frequency selectivity of the subband. A smaller σs corresponds to a narrower

frequency band. The kernel length K is dynamically determined as a proportion

α of the input signal length T, thereby adapting to the analytical requirements

of different time-window settings. In the implementation, fs and σs are treated

as learnable parameters and updated by backpropagation together with the

downstream network. To ensure physiological plausibility and numerical stability,

fs and σs are constrained within the predefined ranges after each parameter update.

In Section 2.2.2, we expanded the initialization description of the G-AFB layer:

Specifically,-the E——EGanalysis range was 3–50 Hz, and the first four harmonics of the stimulus frequencies were considered, i.e., the harmonic order was set to

n = 1, 2, 3, 4. In this study, the number of Gabor subbands was set to S = 4,

corresponding to the first four harmonic ranges of the stimulus frequencies. For

the n-th harmonic band, the initial center frequency was set to the theoretical

center of the corresponding harmonic range, (fmin × n + fmax × n)/2, where fmin

and fmax denote the minimum and maximum stimulus frequencies, respectively.

Harmonic components outside the predefined EEG analysis range were excluded

from the target-related frequency-band construction. Therefore, each learnable

Gabor kernel was initialized around one target-related harmonic band, and its

center frequency and bandwidth were further optimized during training through

end-to-end-backpropagation.

In Section 2.2.2, we also clarified the composition of the prior regularization term:

The final prior regularization loss is obtained as a weighted sum of the three components above:

L prior = λfreqL freq+ λbwLbw+ λdiffL diff (8)

where λfreq, λbw, and λdiff control the relative contributions of frequency alignment,

bandwidth constraint, and bandwidth-difference regularization, respectively. By

using L total as the optimization objective, the parameters of the G-AFB layer can be

jointly optimized with those of the downstream decoding network in an end-to-end

manner.

After Algorithm 1, we added an implementation explanation of the G-AFB for- ward propagation and gradient update:

learnable parameters and then applied to the input EEG signals through depthwise

convolution. Because the Gabor-kernel generation, convolution operation, and

downstream classifier are included in the same computational graph, the gradients

from both the classification loss and the prior regularization loss can be

jointly backpropagated to the G-AFB parameters and the downstream network

parameters. Therefore, the learned center frequencies and bandwidths are not fixed

preprocessing parameters, but are optimized together with the decoding network

during training.

In Section 2.3 Three-Stage Semi-Supervised Domain Adaptation Network (TriS- DANN), we clarified which modules are updated in each training stage:

The trainable modules-diffe——racro—ssthe thre—estages. In—Stage 1-,the feature extractor F and label classifier Cy are optimized using labeled source-domain data,

while the domain classifier Cd is not used. In Stage 2, the label classifier Cy is frozen,

and adversarial training is performed using labeled source-domain samples and

unlabeled target-domain samples. In this stage, the feature extractor F and domain

classifier C d are updated through the gradient reversal mechanism, so that F learns

domain-invariant representations while Cd learns to distinguish source and target

domains. In Stage 3, the domain classifier Cd is no longer used, and the feature

extractor F and label classifier Cy are fine-tuned using the small number of labeled

target-domain calibration samples.

In Section 3.3.1 Experimental Paradigm and Baseline Strategies, we clarified the use of calibration, validation, and test data in TriS-DANN:

fine-tuning stages. In Stage 2, these calibration trials were used without labels

as unlabeled target-domain samples for domain alignment. In Stage 3, the same

calibration trials were used with labels for supervised fine-tuning. The 21 validation

trials were used only for model selection and early stopping, and the 70 test trials

were used-only for final performance evaluation.

In Section 3.3.1, we further specified the model-selection criterion:

semi-supervised domain adaptation procedure: Stage 2 first performs unsupervised

adaptation using the 21 calibration trials without labels, and Stage 3 then fine-tunes

the model using the same 21 trials with labels. Model selection was based on the

target-domain validation accuracy. During training, the checkpoint with the highest

validation accuracy was saved as the final adapted model. The independent test set

was not used for training, domain adaptation, fine-tuning, hyperparameter tuning,

early stopping, or checkpoint selection. The hyperparameters for the three stages of

cross-subject transfer learning are listed-i——nTable—5.

We added the following statement in the Data and Code Availability Statement section of the revised manuscript:

dataset used in this study is available from the Tsinghua BCI Lab download

page (https://bci.med.tsinghua.edu.cn/download.html). The source code of

the proposed framework, including the implementation of G-AFB-tCNN and

TriS-DANN, related training scripts, evaluation scripts, and configuration files, may

be made available from the corresponding author upon reasonable request for

academic and reproducibility purposes. The in-house EEG dataset is not publicly

released because it was collected as part of an ongoing research project and contains

human-subject EEG recordings. Access to the in-house data may be provided upon

reasonable request, subject to institutional approval, project-related restrictions, and

participant-privacy protection requirements.

Comment:

2.4 The figures lack sufficiently detailed legends and explanations.

Mostfigures are not self-contained and lack detailed legends. This reduces readability and makes it difficult to understand the experimental protocol and data flow. This is especially problematic for Figure 1. In Step 2, the figure shows N-1 subjects as the source domain and subject N as the target domain, but it does not clearly distinguish which part of the target data is used for unsupervised adaptation, which part is used for fine-tuning, which part is used for validation, and which part is reserved for testing. If the figure is read without the main text, it may appear that the target subject’s data are used without proper separation, raising concerns about data leakage.
The authors should revise allfigure legends to make thefigures understandable without relying heavily on the main text. In particular, Figure 1 should explicitly indicate the source-domain data, target-domain unlabeled adaptation data, labeled calibration data, validation data, and final test data.

Reply:

We thank the reviewer for this helpful suggestion. We agree that the figures, especially Figure 1, should be more self-contained and should clearly indicate how the data are used in each stage of the proposed framework.

To address this issue, we have revised Figure 1 and its legend to explicitly distin- guish the source-domain data, target-domain unlabeled calibration data for unsuper- vised adaptation, labeled calibration data for supervised fine-tuning, validation data for model selection and early stopping, and final test data for independent evalua- tion. We have also clarified that these target-domain subsets are non-overlapping.

In addition, we checked the legends of the other figures and revised them where necessary by adding brief explanations of the experimental setting, model compo- nents, data flow, plotted quantities, or comparison conditions.

Changes:

In response to this comment, we revised Figure 1 and the legends of all figures to improve their self-contained readability. The revised Figure 1 and figure legends are listed below:

Figure 1: Overall workflow of the proposed framework. The target-domain data are divided into non-overlapping subsets for unsupervised adaptation, supervised fine- tuning, validation, and final testing.

Figure1. Overall workflow of the proposed framework. The target-domain data are divided into non-overlapping subsets for unsupervised adaptation, super- vised fine-tuning, validation, and final testing.
Figure2. Workflow and architecture of the G-AFB layer. The learnable Gabor kernels extract harmonic-related subband features from multichannel EEG sig- nals and feed the resulting features into the downstream decoding network.
Figure3. Impedance map of the selected eight-channel EEG acquisition montage used in the in-house experiment.
Figure4. Experimental procedure of the in-house SSVEP task, including visual cue, stimulation period, and rest period.

30 / 37

Comment:

2.5 Line 177: “a set of shared learnable Gabor kernels” is unclear. Please clarify what “a set” refers to. How many Gabor kernels are included? What does “shared” mean here? Are the kernels shared across channels, trials, subjects, or models?

Reply:

We thank the reviewer for pointing out this unclear wording. We agree that the original phrase “a set of shared learnable Gabor kernels” did not sufficiently explain the number of kernels or the meaning of “shared”. We have revised Section 2.2.1 to clarify that the G-AFB layer contains S = 4 learnable Gabor kernels in this study, corresponding to the first four harmonic-related subbands of the SSVEP stimulus frequencies.

We have also clarified that “shared” means that, within the same trained model, the same set of Gabor-kernel parameters is applied across EEG channels and trials, rather than learning separate kernels for each channel or each trial. These parameters are optimized during training and may differ across independently trained subjects, folds, or model instances.

Changes:

subband feature extraction using S shared learnable Gabor kernels. In this study,

S = 4, corresponding to the first four harmonic-related subbands of the SSVEP

stimulus frequencies. Here, “shared” means that the same set of Gabor-kernel

parameters is applied across EEG channels and trials within the same model,

rather than learning separate Gabor kernels for each channel or each trial. The

Gabor-kernel parameters are optimized during model training and may differ across

independently trained subjects, folds, or model instances.

Comment:

2.6 Line 181: The selection or initialization offs should be explained. The manuscript states that fs is the center frequency of the Gabor kernel, but it is not sufficiently clear how fs is initialized, constrained, or updated during training.

Reply:

We thank the reviewer for this helpful comment. We have revised Section 2.2.1 and Section 2.2.2 to explain the initialization, constraint, and update of fs more clearly. In the revised manuscript, fs denotes the learnable center frequency of the s-th Gabor kernel. Its initial value is determined according to the stimulus-frequency range and the corresponding harmonic bands, rather than being randomly initial- ized.

Specifically, the EEG analysis range is 3–50 Hz, and the first four harmonics of the stimulus frequencies are considered. The initial center frequency of each Gabor ker-

nel is set to the theoretical center of the corresponding harmonic range. During train- ing, fs is treated as a learnable parameter and updated by backpropagation together with the downstream network. To ensure physiological plausibility and numerical stability, fs is constrained within the predefined EEG analysis range after each pa- rameter update.

Changes:

In Section 2.2.1 Gabor Kernels and Adaptive Mechanism, we clarified how fs is constrained and updated during training: