Brain-Inspired Capture: Evidence-Driven Neuromimetic Perceptual Simulation for Visual Decoding

Feixue Shao, Guangze Shi, Xueyu Liu, Yongfei Wu, Mingqiang Wei, Jianan Zhang, Jianbo Lu, Guiying Yan, Weihua YangarXiv 2026 · DOI · arXiv

TL;DR

This paper introduces "Brain-Inspired Capture" (BI-Cap), a novel framework that bridges the gap between brain signals (EEG/MEG) and visual information by simulating the human visual system's processing. BI-Cap achieves state-of-the-art performance in zero-shot brain-to-image retrieval by using biologically inspired transformations and an evidence-driven approach to handle neural signal uncertainty.

What Problem Does This Paper Address?

Decoding visual information directly from neurophysiological signals like EEG and MEG is a significant challenge for brain-computer interfaces (BCIs) and computational neuroscience. Current methods often struggle because of the inherent "systematic" and "stochastic" gaps between how the brain processes information and how visual data is represented computationally. These gaps arise from differences in how neural signals are generated and processed compared to how images are structured and analyzed by algorithms. Crucially, existing approaches often overlook the complex, biologically plausible mechanisms of the Human Visual System (HVS), which could provide a more natural and effective way to align these two modalities.

Methods

BI-Cap tackles this problem through two main innovations:

  1. Neuromimetic Perceptual Simulation (NPS): This paradigm aims to align neural and visual modalities by emulating HVS processing. It involves four biologically plausible transformations applied to input images:
  2. Cortical Magnification Effect: Simulates how the visual cortex dedicates more resources to the foveal region.
  3. Neural Volume Conduction Effect: Introduces additive Gaussian noise to mimic signal spread.
  4. Low-Frequency Dominance: Filters out high-frequency details, reflecting early visual processing's focus on coarse structures.
  5. Retinal Mosaic Sampling: Mimics the discrete sampling of light by photoreceptors. These transformations are fed into a pre-trained image encoder (like CLIP) to generate neuromimetic visual features.

  6. Evidence-Driven Latent Space Representation: To handle the inherent non-stationarity and heterogeneity of neural signals, this approach models uncertainty.

  7. Evidence Weighting Module (EWM): Adaptively combines features from the four neuromimetic transformations by quantifying their reliability using Subjective Logic. Transformations with higher evidence (lower uncertainty) are given more weight.
  8. Linear Weighting Module: A parallel module that captures contextual dependencies using attention.
  9. Feature Purification Module: A bottleneck architecture that enforces feature compactness and reduces noise. This combined approach creates a robust shared latent space for aligning neural and visual representations.

Additionally, MI-Guided Dynamic Blur Regulation is used to optimize the neuromimetic transformations. It dynamically adjusts the blur level based on the alignment quality between neural and visual features, aiming to maximize Mutual Information (MI) and adaptively control visual degradation.

Datasets

The paper evaluates BI-Cap on two public benchmarks:

  • THINGS-EEG2: A large-scale dataset with 64-channel EEG recordings from 10 subjects, based on the THINGS object concept database. It includes training and testing on both intra-subject (within a single subject) and inter-subject (across different subjects) settings.
  • THINGS-MEG: A similar dataset using high-density MEG recordings from 4 subjects, also based on the THINGS object concept database.

Key Insights and Ideas

  • Biologically Plausible Transformations are Key: Directly mimicking aspects of the Human Visual System (HVS) significantly improves the alignment between brain signals and visual representations, suggesting that understanding how we see is crucial for decoding what we see from brain activity.
  • Modeling Uncertainty is Essential: Neural signals are inherently noisy and variable. An "evidence-driven" approach that explicitly quantifies and accounts for this uncertainty leads to more robust and reliable representations.
  • Adaptive Processing Matters: Dynamically adjusting visual processing (like blur) based on the quality of the neural signal alignment, rather than using fixed parameters, leads to better performance. This mirrors how the brain adapts its processing.
  • CNNs Outperform Transformers for EEG: For this specific task, CNN-based image encoders like ResNet-50 showed better performance than Transformer-based models, suggesting that their inductive biases are more suitable for capturing the hierarchical nature of visual information relevant to EEG decoding.

Key Results

BI-Cap achieves state-of-the-art performance in zero-shot brain-to-image retrieval on both THINGS-EEG2 and THINGS-MEG datasets.

  • THINGS-EEG2: Achieved a Top-1 accuracy of 72.4% and Top-5 accuracy of 95.2% in the intra-subject setting, outperforming previous methods by 9.2%.
  • THINGS-MEG: Achieved a Top-1 accuracy of 40.4% and Top-5 accuracy of 69.0%, outperforming previous methods by 8.0%.
  • The method demonstrates strong robustness across different retrieval set sizes and is less sensitive to hyperparameter choices compared to prior methods.
  • Qualitative analysis shows that BI-Cap successfully maps decoded EEG features to their corresponding ground-truth images and retrieves semantically similar images.

Limitations and Future Work

  • Inter-Subject Generalization Gap: While BI-Cap performs well, a significant gap remains between intra-subject and inter-subject decoding. This is attributed to inherent individual variability in neuroanatomy, electrode impedance, and cognitive styles.
  • Transition to Reconstruction: The authors plan to move beyond retrieval to visual reconstruction. By leveraging the learned evidence-driven latent representations, they aim to condition generative models to reconstruct pixel-level images directly from neural signals.
  • Universal Neural Decoder: Achieving a truly universal neural decoder that generalizes seamlessly across diverse populations remains an ongoing challenge due to the fundamental individual differences mentioned above.

Summary updated

Abstract

Visual decoding of neurophysiological signals is a critical challenge for brain-computer interfaces (BCIs) and computational neuroscience. However, current approaches are often constrained by the systematic and stochastic gaps between neural and visual modalities, largely neglecting the intrinsic computational mechanisms of the Human Visual System (HVS). To address this, we propose Brain-Inspired Capture (BI-Cap), a neuromimetic perceptual simulation paradigm that aligns these modalities by emulating HVS processing. Specifically, we construct a neuromimetic pipeline comprising four biologically plausible dynamic and static transformations, coupled with Mutual Information (MI)-guided dynamic blur regulation to simulate adaptive visual processing. Furthermore, to mitigate the inherent non-stationarity of neural activity, we introduce an evidence-driven latent space representation. This formulation explicitly models uncertainty, thereby ensuring robust neural embeddings. Extensive evaluations on zero-shot brain-to-image retrieval across two public benchmarks demonstrate that BI-Cap substantially outperforms state-of-the-art methods, achieving relative gains of 9.2\% and 8.0\%, respectively. We have released the source code on GitHub through the link https://github.com/flysnow1024/BI-Cap.
BibTeX
@misc{shao2026brain,
  author = {Feixue Shao and Guangze Shi and Xueyu Liu and Yongfei Wu and Mingqiang Wei and Jianan Zhang and Jianbo Lu and Guiying Yan and Weihua Yang},
  title = {{Brain-Inspired Capture: Evidence-Driven Neuromimetic Perceptual Simulation for Visual Decoding}},
  year = {2026},
  eprint = {2604.17927},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.17927},
}
Figure