Musical Intelligence — Cycle 17 cross-subject per-piece BOLD encoding artifacts (ds003720)
收藏资源简介:
Reproducibility bundle for Musical Intelligence v9.5.7 Cycle 17 cross-subject per-piece BOLD encoding analysis on ds003720 (Nakai 2021, 4 subjects, 720 music clips). Contains MI 26-D region activation map features, baseline encoder features (MERT-768, CLAP-music-512, Random-26/768), per-subject HRF-lagged per-clip BOLD vectors, and Python analysis scripts for shuffle-null inference, Ridge 5-fold CV encoding, and CKA architecture tests. Reproduces every numerical claim in the paper's §Results "Cross-subject per-piece encoding on ds003720" section without requiring access to the MI engine source code. The four analyses included:- Deney 1: Shuffle-clip null (500 permutations per encoder per subject, Benjamini-Hochberg FDR over 24 tests).- Deney 2: Ridge regression 5-fold CV with nested inner alpha selection, top-5% voxel held-out Pearson r, Fisher-z mean across subjects.- Deney 3: Feature-level CKA (MI-full vs MI-naive, MERT, CLAP, Random-26/768).- Deney 3b: CKA vs per-subject BOLD (encoding-geometry test). Headline results reproducible from this bundle: MI 4/4 subjects pass BH-FDR shuffle-null (effect +0.047); MI-naive 1/4; Random 0/4; MERT 4/4 (+0.054). Ridge held-out r: MI 0.165 vs MI-naive 0.084 (+93% architectural lift); vs Random-26 0.090 (+83% dimensionality-matched lift); MERT 0.221 at 30x more feature dimensions. CKA feature-level (MI-full, MI-naive) = 0.994 (literal R5 H2 floor Δ ≥ 0.02 fails honestly; dimensionality-confounded). Engine IP note: MI engine source code remains under SRC9 Sonic Intelligence LLC copyright. This deposit is an artifact-only release — the numerical encoding analysis is fully reproducible from the deposited features + BOLD + scripts without engine access. Research-use source access is available upon email request to amace@bu.edu (see bundle DOWNLOAD.md). Paper-time engine state is pinned to git tag v9.5.3-paper-time-snapshot (SHA 5beaee5c) in the private SRC9 Musical Intelligence repository. This reproducibility model follows NeurIPS artifact-track guidelines and industry-research conventions (e.g., Gemini tech reports, GPT-4 technical report). Runtime: ~40 min CPU (Apple M2 or equivalent). No GPU required. Bundle integrity verified via MANIFEST.sha256 (38 files).
本套件为针对ds003720(Nakai等人2021年研究,共4名被试、720段音乐片段)的第17周期音乐智能(Musical Intelligence, MI)v9.5.7版本跨被试逐片段血氧水平依赖(Blood Oxygen Level Dependent, BOLD)编码分析的可复现性工具包。其中包含26维脑区激活图特征(MI 26-D)、基准编码器特征(MERT-768、CLAP-music-512、Random-26/768)、逐被试滞后于血氧响应函数(Hemodynamic Response Function, HRF)的逐片段BOLD向量,以及用于置换零假设推断、岭回归5折交叉验证编码与中心化核对准(Centered Kernel Alignment, CKA)架构测试的Python分析脚本。无需获取MI引擎源代码,即可复现论文"ds003720跨被试逐片段编码"章节§结果部分的所有数值结论。 本套件包含四项分析任务: - 分析1:片段置换零假设检验(每位被试对应每个编码器执行500次置换,对24项检验执行本贾尼-霍赫贝格错误发现率(Benjamini-Hochberg False Discovery Rate, FDR)校正)。 - 分析2:带嵌套内部alpha选择的岭回归5折交叉验证,选取前5%体素的保留集皮尔逊相关系数(Pearson r),并计算被试间的费舍z变换均值。 - 分析3:特征级中心化核对准(CKA)检验(对比全量MI模型与朴素MI模型、MERT、CLAP、Random-26/768的特征)。 - 分析3b:针对逐被试BOLD数据的CKA检验(编码几何测试)。 本套件可复现的核心结果如下:4/4名被试的MI模型通过本贾尼-霍赫贝格FDR校正的置换零假设检验(效应量+0.047);朴素MI模型为1/4;Random模型为0/4;MERT模型为4/4(效应量+0.054)。岭回归保留集相关系数:MI模型为0.165,相较于朴素MI模型的0.084提升93%(架构层面增益);相较于Random-26模型的0.090提升83%(维度匹配增益);MERT模型的相关系数为0.221,但其特征维度是前述模型的30倍。特征级CKA检验(全量MI模型与朴素MI模型)结果为0.994(严格R5 H2下界差≥0.02的假设不成立,存在维度混淆问题)。 引擎知识产权说明:MI引擎源代码受SRC9 Sonic Intelligence LLC版权保护。本套件仅发布分析工件——仅通过上传的特征文件、BOLD数据与分析脚本,即可完整复现数值编码分析结果,无需获取MI引擎源代码。如需获取研究用途的源代码,可通过邮件联系amace@bu.edu(详见套件内DOWNLOAD.md文件)。论文提交时的引擎状态已固定至私有SRC9 Musical Intelligence仓库中的git标签v9.5.3-paper-time-snapshot(SHA哈希值5beaee5c)。 本可复现工具包遵循神经信息处理系统大会(Neural Information Processing Systems, NeurIPS)工件赛道指南与工业界-学术界研究惯例(如Gemini技术报告、GPT-4技术报告)。 运行时长:约40分钟CPU运算(Apple M2或等效处理器),无需GPU。套件完整性通过MANIFEST.sha256文件验证(共包含38个文件)。



