Research map / Vision-language and multimodal / Human and embodied modeling

Neural‑Visual Decoding via Cognitive‑guided Adaptive Blurring and Information‑Constrained Alignment

NeurIPS 2026 · Vision-language and multimodal › Human and embodied modeling · approach: fMRI visual decoding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CAIA, which fuses center-biased and saliency-guided blurred images via EEG-driven cross-modal attention, adds an information-bottleneck wavelet band-selection encoder and a distribution-aware boundary loss, and reports state-of-the-art zero-shot EEG/MEG-to-image retrieval on THINGS.

Problem

Zero-shot EEG-based visual decoding (brain-to-image retrieval) is hampered by a granularity mismatch between fine visual detail and coarse EEG information, and by the low SNR of EEG signals.

Approach

Bidirectional modulation: adaptively blur images using a cross-modal attention fusion of center-bias and saliency-guided blur paths driven by EEG features, while screening EEG frequency bands with a learnable IB-style wavelet selector and calibrating outlier similarities with a Gaussian-interval boundary loss.

Key results

• On THINGS-EEG subject-dependent 200-way zero-shot retrieval, CAIA reaches 70.3% Top-1 and 96.1% Top-5 average, versus 50.9%/79.7% for UBP.
• Subject-independent THINGS-EEG average is 16.4% Top-1 / 45.7% Top-5 (UBP 12.4% / 33.4%).
• On THINGS-MEG subject-dependent, CAIA reaches 31.8% / 62.6% vs UBP 26.7% / 55.2%.

Surprising findings

• The boundary calibration loss alone slightly degrades performance (52.4% vs 54.5% Top-1), which the authors attribute to it acting as an outlier detector rather than a corrector.

Limitations

• Cognitive priors (center bias, saliency) presuppose a paradigm with sustained central fixation and may degrade under free viewing.
• Inference latency rises from 5.88 ms to 14.29 ms and parameters from 43.7M to 107.7M relative to UBP.
• Evaluated only on THINGS-EEG and THINGS-MEG, both with RSVP-type paradigms.

Experimental setup

Datasets: THINGS-EEG THINGS-MEG

Benchmarks: THINGS-EEG THINGS-MEG

Metrics: Top-1 accuracy Top-5 accuracy

Baselines: BraVL NICE ATM CogCap VE-SDN MB2C Neural-MCRL UBP

Builds on: CLIP UBP NICE ATM

Closest papers by meaning

• MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding (NeurIPS 2026 · similarity 0.91)
• Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion (NeurIPS 2024 · similarity 0.89)
• What Does the Brain See? Multiview Neural Representation to Demystify the Brain-Visual Alignment (NeurIPS 2026 · similarity 0.89)
• Bridging the Vision-Brain Gap with an Uncertainty-Aware Blur Prior (CVPR 2025 · similarity 0.88)
• Human-Aligned Image Models Improve Visual Decoding from the Brain (ICML 2025 · similarity 0.88)
• Brain decoding: toward real-time reconstruction of visual perception (ICLR 2024 · similarity 0.87)