fmri

A large international study published in Nature finds that the three most common measures of cognitive reappraisal capacity—trait questionnaires, task-based affective ratings, and amygdala down-regulation via fMRI—show almost no shared variance, calling into question our ability to measure a key process underlying cognitive behavioral therapy.

该论文比较了扩散式与自回归多模态大语言模型生成的标注与人类在观看卡通短片时的Theory of Mind和情感效价评分的时间一致性，并检验其对fMRI脑反应的预测能力，发现MLLM标注可作为脑预测性社会情感内容标注的可扩展工具，但与人类评分存在系统性差异且对齐效果因模型和影片而异。

发布了一个名为 Cajal-1 的 2.18 亿参数全脑皮层编码模型，基于 2000 人观看影片、电视和图片时的 fMRI 数据训练，可预测视觉、听觉和语言刺激在 20,484 个皮层位置上的响应。模型与 Colab 演示已在 Hugging Face 上开源。

Researchers at the Weizmann Institute have developed an AI tool that reconstructs what a person is looking at from high-resolution fMRI brain scans with remarkable precision, and can also predict brain activity from images. Experts praise its potential to help locked-in patients and understand dreams, while warning it could enable surreptitious extraction of inner thoughts.

This arXiv paper builds expert LLM variants specialized for six cognitive domains via prompting and fine-tuning, showing each expert's representations align better with the corresponding specialized brain system across three base models and three fMRI datasets.

FRIST is a two-stage EEG decoding framework that leverages fMRI data to improve EEG-only individual-finger BCI decoding, demonstrating increased accuracy in movement execution and motor imagery tasks.

A new fMRI study in iScience finds that dogs process happy human facial expressions in brain regions linked to higher cognition and reward, and can distinguish between negative expressions like anger, sadness, and fear.

This paper proposes MPP-GNN, a Meta Probabilistic Pooling GNN that adaptively discovers subject-specific brain modules for fMRI-based Alzheimer's disease classification, achieving state-of-the-art AUC on two public datasets and aligning with canonical brain network organization.

This study demonstrates that contextual semantic relevance, measuring how strongly an incoming word relates to its recent semantic context, reliably predicts fMRI BOLD responses during naturalistic speech comprehension across two datasets, whereas surprisal (local probabilistic expectation) does not. The findings support that slow hemodynamic responses are especially sensitive to contextual semantic integration rather than local prediction.

This paper improves the Huth encoding pipeline for fMRI decoding and introduces fMRIFlamingo, a direct fMRI-to-text framework using Llama 3.2, but finds that decoding success is driven by the language prior rather than neural input.

This paper introduces Augmented Sparse Encoding Models to interpret brain responses to language using sparse features from language models, validated on high-field 7T fMRI data. It recovers known neural tuning properties and discovers a new voxel population tuned to people-related content.

This paper tracks how different learning rules (backprop, feedback alignment, predictive coding, STDP) affect the alignment of CNN representations with human fMRI across training. It finds that backprop destroys V1 alignment in one epoch, while local rules preserve it, suggesting a trade-off between building higher-level representations and retaining early visual features.

CortexMAE is a family of fMRI foundation models trained on 2.1K hours of open fMRI data, accepted to ICML. The release also includes Brainmarks, an open benchmark suite for evaluating such models.

This paper introduces BrainSimSiam, a lightweight self-supervised framework using siamese networks to learn robust fMRI representations from positive-only pairs, achieving strong performance on downstream tasks even with limited data.

Brain-IT-VQA framework decodes visual content from fMRI signals using transformer architecture, outperforming previous methods. The authors also introduce NSD-VQA, a new dataset with richer annotations for evaluating fMRI-based visual question answering.

Proposes FM-fMRI, an event-conditioned flow matching model that synthesizes task fMRI time series from resting-state fMRI, achieving superior spectral and connectivity agreement over baselines on the Human Connectome Project and an internal autism cohort, and improving downstream autism classification.

This paper investigates brain-LLM alignment across English, Chinese, and French using fMRI data and multiple LLMs, finding that training-language dominance and typological distance, not an inherent English advantage, drive alignment patterns.

Google releases MoGen model, using synthetic neurons to reduce connectome reconstruction error rate by 4.4%, saving 157 person-years of annotation time; meanwhile, AI can already decode thoughts from fMRI, and the proliferation of portable brain sensors raises brain privacy concerns.

BrainCause framework uses generative and brain models to identify causal neural representations in the human brain, demonstrating that activation alone is insufficient for confirming concept representation.

This paper demonstrates that fine-tuning language encoding models on fMRI data improves their ability to predict neural activity from ECoG recordings, despite fMRI's lower temporal resolution. The findings suggest that abundant 'slow' fMRI data can enhance models for 'fast' ECoG data.