acceptodds
Withdrawn from ICLR 2027

CFBA: Contour-preserving Foveated Blur Alignment for Neural Visual Decoding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Decoding visual content from brain signals helps characterize human perception and supports brain–computer interfaces. However, existing approaches to visual supervision adaptation do not explicitly combine contour preservation with spatially varying detail suppression, potentially limiting neural–visual alignment. To address this limitation, we propose Contour-preserving Foveated Blur Alignment (CFBA), which combines explicit preservation of estimated contour pixels with spatially varying suppression of surrounding detail. EGNet predictions are refined through morphological closing and contour-length filtering to construct a contour mask. An exponential radial profile blends the original image with its Gaussian-blurred counterpart, retaining more original detail near the image center. Restoring the masked contour pixels produces the supervision target. A Residual Projection Encoder (RPE) aligns neural recordings with frozen CLIP ResNet-50 features of this target through symmetric contrastive learning. A separate ATM–ViT-H-14 branch learns alignment under original-image supervision. The RPE supports retrieval and classification. For reconstruction, a diffusion prior converts the ATM embedding into the primary condition, while a linear projection trained through the generation objective maps the RPE embedding into an auxiliary condition. IP-Adapters inject both conditions into a pretrained SDXL generator. Extensive experiments on THINGS-EEG and THINGS-MEG demonstrate that CFBA outperforms state-of-the-art methods across retrieval, classification, and reconstruction. The code is available at https://anonymous.4open.science/r/CFBA-D9CD.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Electroencephalogram, Magnetoencephalography, Visual Decoding

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_uYYEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CFBA restores contour pixels within a foveated blur to construct visual supervision for EEG/MEG decoding. A neural encoder aligned to these targets supports retrieval and classification, while its embedding complements an original-image branch for diffusion reconstruction. Experiments evaluate two recording modalities and three tasks.

Critical strengths

Restoring contours improves retrieval over foveated blur alone from 49.7% to 54.5% for EEG and 26.5% to 28.1% for MEG (Table 1), directly supporting the central design choice. Evaluation across modalities and tasks, together with alternative blur and contour-detector comparisons, provides useful evidence beyond one favorable configuration.

Critical weaknesses
  1. The model-selection procedure is not sufficiently specified. The paper studies several target and encoder configurations and uses early stopping, but does not identify the validation partition and criterion used to select the reported models. Confirming that these decisions exclude the unseen test concepts is important for interpreting the gains; the missing description alone is not evidence that test-set selection occurred.
  2. The reconstruction training description leaves a consequential ambiguity. Section 3.3 states that the original-image branch updates ViT-H-14, whereas Figure 2 depicts the CLIP encoder as frozen. Together with the limited account of what is retrained in Table 1's reconstruction variants, this makes it difficult to determine whether those gains isolate the supervision target or also reflect changes to the primary generative condition.
  3. The broad superiority claim is stronger than the comparison establishes. Section 4.1 correctly notes that CFBA leads on five of seven EEG reconstruction metrics, which is narrower than Section 1's claim of surpassing all methods across tasks. Since blur-based neural alignment already appears in UBP, clarifying which baseline results share the encoder, candidate sets, repetition averaging, and generator would better isolate CFBA's specific advance.
  4. The experiments do not fully distinguish contour specificity from restoring additional image information. Table 1 establishes that adding contour pixels helps, and Table 3 usefully compares detectors, but all restoration variants prioritize boundaries. A single control restoring the same number of non-contour pixels under otherwise identical blur would test whether the benefit specifically comes from shape-defining boundaries, strengthening the perceptual motivation without requiring another dataset.
  5. The alignment analysis does not independently establish the proposed explanation. Table 2 measures similarity and separation after training neural encoders toward different visual targets, whose feature geometry also changes with the transformation. These are useful decoding diagnostics, but higher scores do not by themselves show that the target better captures information recoverable from neural recordings; a common-reference analysis would help distinguish this explanation from easier target discrimination.
Questions for authors
  1. Which data and metric determine early stopping and target-parameter selection, and were the unseen test concepts excluded from every such decision?
  2. Is ViT-H-14 frozen or updated, what is retrained in the Table 1 reconstruction variants, and which comparisons use matched training and evaluation settings?
  3. Does contour restoration still help against an equal-area non-contour restoration control, and does its alignment advantage persist when evaluated against a common visual reference?
Reviewer_3frnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a modified visual target for EEG/MEG decoding. The images are blurred with central fovea, and then edge information are added back in. These new image targets are then aligned to EEG in the CLIP space.

Critical strengths

Adapting the visual target to what can actually be recovered from EEG is a reasonable direction, and the method is simple to implement.

Critical weaknesses

Overall the novelty on this work is very limited. I find several of the reported numbers very confusing:

  • With 200 test images per subject and 10 subjects, mean accuracies should be multiples of 0.05%. How is it possible to get values like 79.82 (Fig. 4b)?

  • The 900–1000 ms window has chance-level retrieval but gets SSIM 0.492 and CLIP 0.740 (Table 9). Frontal channels get higher SSIM than occipital-parietal (Table 13). Are the reconstruction metrics actually measuring anything about the neural data? A shuffled EEG-image baseline could be informative.

  • Is there a separate validation for tuning the hyperparameters?

  • Section A.2 claims the representation transfers to category matching. But classification is 7.45% compared to 54.5% retrieval on the same 200 categories. This suggests the embedding captures visual similarity more than semantics. On this note, the selected images are very cherrypicked.

  • In Fig. 2(a), the final target looks sharper than I_F, but Eq. 6 only restores thin contour pixels. Is this representative of the actual targets?

Fig. 3 has no scale or normalization described.

Questions for authors

See above.

Reviewer_F6HVOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Contour-preserving Foveated Blur Alignment (CFBA) for visual decoding from EEG and MEG signals. The central idea is to modify the visual supervision targets by preserving salient object contours while progressively suppressing peripheral image details. The method is evaluated on THINGS-EEG and THINGS-MEG across image retrieval, zero-shot classification, and reconstruction. The reported results are competitive, and the ablation studies generally support the proposed combination of contour preservation and foveated smoothing.

Critical strengths
  • Adapting the visual targets is a meaningful direction for neural-visual representation learning.
  • CFBA improves retrieval on both EEG and MEG.
Critical weaknesses
  • The central motivation of this paper closely overlaps with prior fMRI studies on reducing semantic mismatch between neural representations and visual supervision. For example, Mind-SA (ICCV 2025) [1], and Yo'Mind (ICLR 2026) [2] have already investigated the discrepancy between visual information captured by pretrained models and the semantic content selectively encoded in brain activity. Both approaches address this issue by adapting visual supervision to the information represented in neural responses. These studies are highly relevant to the motivation of CFBA, yet neither is discussed in the related work.

[1] Bridging the gap between brain and machine in interpreting visual semantics: Towards self-adaptive brain-to-text decoding

[2] Reducing Semantic Mismatch in Brain-to-Text Decoding Through Personalized Multimodal Masking

  • The main contribution is the combination of an image-centered blur profile with a contour mask obtained from a pretrained EGNet model. Foveated smoothing has already been explored in closely related work, particularly UBP. Although preserving contours is a sensible extension, the proposed transformation is largely deterministic and does not depend on the neural recordings or learned cross-modal correspondence. As a result, it is unclear whether the work establishes a more general principle for alignment or primarily demonstrates the effectiveness of a particular image preprocessing strategy.

  • The method is motivated by contour organization and the spatially varying resolution of human vision. However, the protected contours are derived from an off-the-shelf salient object detector and are not necessarily equivalent to the boundaries represented in human visual cortex.

  • The reconstruction pipeline combines two independently trained branches, different visual backbones, a diffusion prior, and separate adapter pathways. Table 4 shows that the asymmetric assignment of ATM–ViT-H-14 as the primary condition and RPE–RN50 as the auxiliary condition performs best. However, this comparison does not fully establish which properties of the auxiliary representation account for the improvement. I would be particularly interested in a controlled comparison where the primary branch and generation pipeline remain fixed, while the auxiliary branch is trained with original-image, foveated-blur, and CFBA supervision under otherwise identical settings.

Questions for authors
  • How sensitive is CFBA to the spatial location of the target object? In particular, does the proposed center-weighted foveation remain effective for off-center objects?

  • Does CFBA consistently improve performance over original-image and foveated-blur supervision when the neural encoder and CLIP image backbone are varied independently?

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

Related papers

Discussion (0)

Sign in to comment.