acceptodds
Under review as a conference paper at ICLR 2027

BioBridge: Bridging Brain and Vision Representations with Biological Inductive Biases

Abstract

EEG-based visual neural decoding seeks to infer visual content from neural responses and supports applications such as image retrieval. However, progress is limited by the scarcity of paired EEG and image data and by the substantial mismatch between high-fidelity digital images and biological visual representations shaped by retinotopic organization and subject-dependent neuroanatomy. To address these challenges, we propose BioBridge, a biologically guided brain–vision alignment framework that incorporates structured physiological inductive biases into representation learning. Specifically, Adaptive Blur with Visual Priors uses retinotopic weighting to transform visual inputs into perceptually structured representations, thereby reducing the discrepancy between digital stimuli and biological perception. Biomimetic Visual Feature Extraction then learns complementary visual representations at multiple levels that reflect hierarchical cortical processing and improve robustness to subject variability. Finally, Multi-Level Bidirectional Contrastive Learning jointly aligns EEG and visual representations at different semantic levels through symmetric contrastive objectives in a shared embedding space. Experiments demonstrate that BioBridge achieves 80.5% Top-1 accuracy and 97.6% Top-5 accuracy in zero-shot retrieval from EEG to images, substantially outperforming previous methods and exhibiting strong generalization across subjects and experimental settings. Code is available at https://anonymous.4open.science/r/BioBridge-227/.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neuroscience Applications, Cognitive Science, Neural Decoding, Multimodal Learning, Zero-shot Learning

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

BioBridge injects three biological inductive biases into contrastive brain–vision alignment: (i) Adaptive Blur with Visual Priors treats the image center as the fovea and fuses a Gaussian-blurred copy with the original via a learnable eccentricity-dependent weight; (ii) Biomimetic Visual Feature Extraction splits EEG channels into low- and high-level streams with channel attention initialized from electrode positions, then derives high-level features by cross-attention conditioned on low-level ones; (iii) Multi-level Bidirectional Contrastive Learning decouples image features into low- and high-level and aligns them with EEG features via symmetric InfoNCE in both directions. At inference, retrieval uses the EEG encoder alone. The method reports 80.5% top-1 and 97.6% top-5 intra-subject on THINGS-EEG, versus 50.9% and 79.7% for the strongest compared baseline.

Critical strengths

Evaluation spans two neural modalities (THINGS-EEG with 10 subjects and THINGS-MEG with 4 subjects), 4 EEG encoders × 4 pretrained visual encoders, both intra- and inter-subject (leave-one-subject-out) settings, and a large set of ablations covering each module, alternative blur priors, retrieval pool size K, top-K metrics, and subject variability. The reported improvement over the strongest compared baseline is substantial (80.5% vs. 50.9% top-1), and anonymous code is released, which supports reproducibility.

Critical weaknesses
  1. For ABVP, the foveated blur prior is already the core of UBP (CVPR 2025), which fuses a Gaussian-blurred image with the original using an eccentricity-dependent weight α(i,j)=exp(−λ·d/L) under the same RSVP fixation assumption, and which already reports an ablation comparing uniform blur versus fovea blur; the present paper's change is essentially replacing a fixed exponential decay with a learnable logistic gate. Related foveated degradation is also used in BVSA (CVPR 2026). For BVFE, the idea of hierarchical low-level-to-high-level integration via progressive cross-attention, together with hierarchical contrastive alignment to CLIP features, is already the main design of ViEEG. For MBCL, symmetric bidirectional InfoNCE is the standard CLIP objective, bidirectional EEG–image alignment is explicitly claimed by NeuroBridge (AAAI 2026), and multi-level/hierarchical alignment is standard in vision–language pretraining (e.g., PyramidCLIP).

  2. Tables 1–2 compare only against methods up to 2025 (ending at UBP, 50.9%). Several direct competitors from 2026 are absent and, in some cases, are only mentioned in passing in the related work: NeuroBridge (AAAI 2026, 63.2% top-1 on THINGS-EEG), BVSA (CVPR 2026, 58.2%), the multiview EEG representation model (2026, 54.8%), BSG-DPM (2026, 47.4%), and CI-BVCL (2026, 42.5%), as well as ViEEG (40.9%). Because the strongest published number excluded from the tables (63.2%) is well above every baseline the paper does compare against, the reported margin over "previous state of the art" cannot be assessed without these comparisons.

Questions for authors
  1. UBP (CVPR 2025) already fuses a Gaussian-blurred image with the original using an eccentricity-dependent weight and reports a uniform-blur vs. fovea-blur ablation. What is the incremental gain of your learnable logistic gate on top of UBP's fovea blur under the same training setup? Without this ablation, the novelty of ABVP cannot be judged.
  2. Could you report numbers for NeuroBridge (AAAI 2026, 63.2% top-1) and BVSA (CVPR 2026, 58.2%) under your protocol? These are direct competitors and are currently excluded from Tables 1–2.
Reviewer_1k7MOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BioBridge combines learnable foveated blurring, hierarchical EEG encoding with physiological channel priors, and multi-level bidirectional contrastive learning for zero-shot brain-to-image retrieval. It reports 80.5% within-subject Top-1 accuracy on 200-way THINGS-EEG retrieval using trial-averaged EEG, with additional evaluations across subjects and on MEG data.

Critical strengths

The reported improvement over the listed baselines is substantial. The experiments cover multiple encoders, image transformations, and both EEG and MEG. Table 3 also shows that ABVP and BVFE each improve performance within the full framework.

Critical weaknesses
  1. Missing comparisons with closely related methods. The paper cites NeuroBridge, which learns projections for both modalities, and ViEEG, which uses hierarchical EEG encoding, cross-attention, and hierarchical alignment. Neither is included in the quantitative comparisons. Given their overlap with the proposed design, comparisons under comparable evaluation conditions are needed to establish what BioBridge adds beyond these approaches.

  2. The ablations do not isolate the contribution of the physiological design. In Table 3, removing MBCL also removes trainable image-side modules, while removing BVFE removes both the channel prior and cross-attention. These results support the combined modules but leave open how much the gains depend on physiological priors rather than general representation adaptation and hierarchical processing. Table 4 provides evidence for the visual radial prior. Two further controls would clarify the remaining contribution: a single-level alignment baseline retaining trainable projections for both modalities, and a BVFE variant with the same architecture but uniform or random channel initialization.

Questions for authors

Please refer to the Weakness part.

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

This paper proposes a biologically inspired framework for EEG-based visual decoding that aims to improve brain-vision alignment by incorporating physiological priors. The framework combines adaptive fovea-centered image blurring, hierarchical EEG feature extraction with channel attention, and multi-level bidirectional contrastive learning to align neural and visual representations. Experiments on THINGS-EEG and THINGS-MEG report zero-shot image retrieval performance in both intra-subject and inter-subject settings.

Critical strengths

This paper achieves good zero-shot EEG-to-image retrieval results on THINGS-EEG and THINGS-MEG.

Critical weaknesses

1, The novelty is not sufficiently established. Several key ideas, including biologically inspired visual blurring, hierarchical EEG representations, and multi-level contrastive alignment, have already been used in existing studies such as UBP (CVPR 2025), ViEEG (ICML 2026), and "Leveraging Visual Blur Perception Characteristics for EEG Decoding" (AAAI 2026). In particular, the AAAI 2026 study reports 80.0% Top-1 accuracy on THINGS-EEG, compared with 80.5% achieved by BioBridge. Moreover, this paper does not consistently outperform this closely related work on inter-subject EEG and THINGS-MEG evaluations. The omission of such relevant comparisons weakens the claimed novelty and SOTA performance.

2, The paper claims that ABVP captures individual-specific retinal topography and cortical organization through learnable fovea-centered blurring. However, this proposed module essentially applies a spatially weighted combination of original and Gaussian-blurred images, with parameters optimized using cross-modal contrastive learning. Improved retrieval performance does not establish that the learned blur functions correspond to actual individual retinal structures or cortical processing.

3, The BVFE module is claimed to capture hierarchical visual processing through physiologically initialized EEG channel weights and cross-attention. However, the separation of low- and high-level representations is primarily determined by predefined channel weighting and network architecture, without direct evidence that these features correspond to distinct cortical processing stages. The channel attention visualizations in Fig. 6 only demonstrate different spatial attention patterns and cannot establish genuine cortical functional specialization. Also, the ablation study does not adequately isolate the contribution of physiological channel priors from conventional learnable attention mechanisms.

4, The experimental evidence does not sufficiently isolate the contributions of the proposed biological priors from general image augmentation and conventional contrastive learning. In particular, Gaussian noise combined with ABVP achieves 86.2% Top-1 accuracy, exceeding the default BioBridge configuration, suggesting that generic visual regularization may contribute substantially to the reported improvements. Additionally, the main paper and appendix contain inconsistencies in the reported Top-5 accuracy. The evaluation also relies on averaging 80 repeated EEG trials per test image, while statistical analyses across independent training runs are absent.

Questions for authors

Please refer to the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.