epfl-neuroai/multimodal-brain-scaling
收藏资源简介:
该数据集名为Multimodal Brain Scaling,是一个用于研究视觉皮层任务和数据优化模型的多模态缩放规律的大规模基准数据集。它基于ICML 2026论文《Multimodal Scaling Laws for Task & Data-Optimized Models of Visual Cortex》发布,旨在统一评估模型与大脑对齐性能。数据集涵盖8个神经数据集,包括猕猴电生理学以及人类的功能磁共振成像(fMRI)、脑电图(EEG)和脑磁图(MEG)数据,并评估了超过600个视觉模型。通过三个缩放轴(预训练资源、神经微调以及从特征到神经响应的映射),数据集揭示了预训练饱和、互补神经微调和映射缩放等趋势。数据以Parquet文件格式存储,包含预训练结果、层搜索结果、微调结果和映射结果等表,并附带元数据如数据集维度、模型信息和噪声天花板。该数据集可用于神经科学、神经AI和计算机视觉研究,帮助优化大脑对齐模型的数据、计算和监督策略。
The dataset, named Multimodal Brain Scaling, is a large-scale benchmark for studying multimodal scaling laws in task- and data-optimized models of the visual cortex. It is released in conjunction with the ICML 2026 paper Multimodal Scaling Laws for Task & Data-Optimized Models of Visual Cortex, providing a unified framework to evaluate model-to-brain alignment. It spans 8 neural datasets, including macaque electrophysiology and human fMRI, EEG, and MEG recordings, and evaluates over 600 vision models under a consistent pipeline. The dataset explores three scaling axes: pretraining resources, neural fine-tuning, and the mapping from features to neural responses, revealing trends such as pretraining saturation, complementary neural fine-tuning, and mapping scaling. Data is stored in Parquet files, containing result tables for pretraining, layer search, finetuning, and mapping, along with metadata like dataset dimensions, model information, and noise ceilings. It supports neuroscience, neuroAI, and computer vision research, offering practical guidance for building brain-aligned models.




