neural-encoding-evaluation-meeg

name	neural-encoding-evaluation-meeg
version	v1.0.0
last_updated	2026-04-18T00:00:00.000Z
description	Evaluation framework for neural encoding models using MEEG (Mutual-information-based Estimation of Encoding model goodness-of-fit). Provides systematic methodology for assessing how well neural models predict brain activity, with information-theoretic metrics and cross-validation protocols.
category	neuroscience
tags	["encoding-models","model-evaluation","mutual-information","neural-data","information-theory","model-selection"]
paper	{"title":"Neural Encoding Model Evaluation (MEEG)","published":"2026-04-17","url":"https://arxiv.org/abs/2604.12463"}
activation	encoding model, model evaluation, neural data, mutual information, model selection, goodness-of-fit

Neural Encoding Model Evaluation (MEEG)

概述

神经编码模型评估框架，使用基于互信息的评估指标（MEEG）系统化地评估神经模型预测脑活动的能力。提供信息论指标和交叉验证协议。

核心问题

神经编码模型（如 pRF 模型、DNN 特征编码）的性能评估缺乏统一标准。需要信息论框架来量化模型对神经数据的解释能力。

方法论

MEEG 指标

def compute_meeg(predicted, observed):
    """计算基于互信息的编码模型拟合优度"""
    # 估计联合分布 p(predicted, observed)
    joint_dist = estimate_joint(predicted, observed)
    # 计算互信息
    mi = mutual_information(joint_dist)
    # 归一化为解释方差当量
    meeg_score = normalize_mi(mi)
    return meeg_score

评估协议

1. 交叉验证：k-fold 交叉验证，避免过拟合
2. 基线比较：与简单基线模型比较
3. 噪声上限：估计数据本身的可预测性上限

模型选择

• 使用 MEEG 分数进行模型比较
• 考虑模型复杂度（AIC/BIC）
• 多模态数据的联合评估

应用场景

• 视觉编码模型：评估 DNN 特征对 V1-V4 活动的预测
• 语言编码模型：评估语言模型对 ECoG 响应的预测
• 多模态编码：联合评估跨模态的编码性能

参考文献

@article{meeg2026,
    title={Neural Encoding Model Evaluation (MEEG)},
    journal={arXiv preprint arXiv:2604.12463},
    year={2026}
}

---

Generated on 2026-04-18