메뉴
EEG基础模型的身份陷阱诊断审计框架。FMScope协议用于诊断EEG基础模型是否陷入身份陷阱（subject identity shortcut），提供五项诊断指标分离真实生物标志物与身份特征。
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/hiyenwong/ai_collection --skill identity-trap-eeg-foundation-models명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
Graph-native Python reimplementation of the Information Dynamics of Music (IDyOM) model that represents predictive memories as explicit graph objects for musical expectation modeling and network analysis.
Physics-aware end-to-end deep reinforcement learning methodology for quadcopter control with actuator dynamics modeling.
Reinforced Dreamer methodology for asymmetric reinforcement learning using latent guidance to improve world model representations and behaviors in model-based RL.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | identity-trap-eeg-foundation-models |
| description | EEG基础模型的身份陷阱诊断审计框架。FMScope协议用于诊断EEG基础模型是否陷入身份陷阱（subject identity shortcut），提供五项诊断指标分离真实生物标志物与身份特征。 |
| version | 1.0.0 |
| author | Jun-You Lin, Ying Choon Wu, Tzyy-Ping Jung |
| arxiv_id | 2606.06647 |
| submission_date | 2026-06-04T00:00:00.000Z |
| categories | ["neuroscience","machine-learning","eeg","foundation-models"] |
| tags | ["identity-trap","eeg-foundation-model","fmscope","shortcut-learning","diagnostic-audit","subject-variance","aperiodic-1f","cross-subject-validation"] |
| activation_keywords | ["身份陷阱","EEG基础模型","shortcut learning","subject identity","frozen representation","FMScope","variance decomposition","aperiodic 1/f","cross-subject marker"] |
| setup_needed | false |
本技能提取自 arXiv:2606.06647 (2026-06-04)，论文标题 "The Identity Trap in EEG Foundation Models: A Diagnostic Audit"。该论文首次发现并诊断了EEG基础模型的"身份陷阱"问题——模型可能通过学习被试身份特征而非真实生物标志物来达到高准确率。
EEG基础模型在临床静息态EEG上报告了高准确率，但在被试不交叉的交叉验证下，高准确率可能是：
论文提出问题：能否在微调前从表示层面诊断身份陷阱？
FMScope 是冻结表示协议，包含五项诊断：
分析表示方差分解：
方法：
# 方差分解
subject_variance = compute_variance_across_subjects(representation)
label_variance = compute_variance_across_labels(representation)
ratio = subject_variance / null_model_variance
诊断标准：
消除被试身份轴：
方法：
# PCA分解
pca = PCA(n_components=k)
components = pca.fit_transform(representation)
# 识别被试相关成分
subject_components = identify_subject_axis(components, subject_ids)
representation_cleaned = representation - subject_components
效果：
移除 aperiodic 1/f 信号：
方法：
# 功率谱分解
spectrum = compute_power_spectrum(eeg_signal)
# 分离 aperiodic 1/f
aperiodic = fit_1_f_slope(spectrum)
# 消融
eeg_cleaned = remove_aperiodic_component(eeg_signal, aperiodic)
发现：
逐层分析标签探测：
方法：
# 层级探针
for layer in model.layers:
probe = train_linear_probe(layer.output, labels)
probe_accuracy[layer] = evaluate_probe(probe)
分析：
分析被试内标签方向一致性：
方法：
# 被试内标签方向
for subject in subjects:
subject_representation = get_subject_rep(subject)
label_direction = compute_label_gradient(subject_representation, labels)
consistency = measure_direction_consistency(label_direction)
诊断：
论文设计了2x2布局的12个实验对：
| Label Varies Within Subject | Label Fixed Within Subject | |
|---|---|---|
| Has Cross-Subject Marker | 3 pairs | 3 pairs |
| No Cross-Subject Marker | 3 pairs | 3 pairs |
身份陷阱是 shortcut learning 的物理实例化：
仅靠被试不交叉验证无法排除身份陷阱：
微调如何放大特征：
冻结表示诊断优势：
提取冻结表示：
进行方差分解：
消除被试轴：
消融 aperiodic 1/f：
层级分析：
评估方向一致性：
生成诊断报告：
评估三个基础模型：
临床 EEG 诊断：
验证 EEG 研究：
身份陷阱揭示 shortcut learning 的物理实例：
暴露传统验证方法的局限性：
建立 EEG 基础模型审计标准：
提出新的表示质量评估范式：
import numpy as np
from sklearn.decomposition import PCA
from sklearn.linear_model import LogisticRegression
class FMScope:
"""EEG 基础模型身份陷阱诊断"""
def __init__(self, model, representations, subject_ids, labels):
self.model = model
self.reps = representations
self.subject_ids = subject_ids
self.labels = labels
def variance_decomposition(self):
"""方差分解诊断"""
# 试方差
subject_groups = self.group_by_subject(self.reps, self.subject_ids)
subject_variance = self.compute_variance(subject_groups)
# 标签方差
label_groups = self.group_by_label(self.reps, self.labels)
label_variance = self.compute_variance(label_groups)
# 随机零模型
null_variance = self.null_model_variance()
# 比率
subject_ratio = subject_variance / null_variance
label_ratio = label_variance / null_variance
return {
'subject_ratio': subject_ratio,
'label_ratio': label_ratio,
'is_identity_trap': subject_ratio > 13
}
def subject_axis_erasure(self):
"""被试轴消除诊断"""
# PCA 分解
pca = PCA(n_components=10)
components = pca.fit_transform(self.reps)
# 识别被试轴
subject_axis = self.identify_subject_axis(components, self.subject_ids)
# 消除被试轴
reps_cleaned = self.reps - subject_axis
# 评估改善
original_acc = self.evaluate_label_decoding(self.reps)
cleaned_acc = self.evaluate_label_decoding(reps_cleaned)
return {
'gain': cleaned_acc - original_acc,
'cleaned_reps': reps_cleaned
}
def aperiodic_ablation(self, eeg_signals):
"""Aperiodic 1/f 消融"""
# 功率谱分解
spectra = [self.compute_spectrum(sig) for sig in eeg_signals]
# 分离 aperiodic
aperiodic = [self.fit_1f_slope(spec) for spec in spectra]
# 消融
signals_cleaned = self.remove_aperiodic(eeg_signals, aperiodic)
# 重新提取表示
reps_cleaned = self.model.extract(signals_cleaned)
# 评估身份探针变化
original_probe = self.subject_probe(self.reps)
cleaned_probe = self.subject_probe(reps_cleaned)
return {
'probe_drop': original_probe - cleaned_probe,
'reps_cleaned': reps_cleaned
}
def layerwise_label_probing(self):
"""层级标签探测"""
layer_accuracies = {}
for layer_idx, layer_output in self.model.get_layer_outputs(self.reps):
probe = LogisticRegression()
probe.fit(layer_output, self.labels)
acc = probe.score(layer_output, self.labels)
layer_accuracies[layer_idx] = acc
return layer_accuracies
def within_subject_direction_consistency(self):
"""被试内方向一致性"""
consistencies = []
for subject in np.unique(self.subject_ids):
subject_reps = self.reps[self.subject_ids == subject]
subject_labels = self.labels[self.subject_ids == subject]
if len(np.unique(subject_labels)) > 1:
# 计算标签梯度
direction = self.compute_label_gradient(subject_reps, subject_labels)
consistency = self.measure_consistency(direction)
consistencies.append(consistency)
return np.mean(consistencies)
def generate_diagnostic_report(self):
"""生成诊断报告"""
variance_result = self.variance_decomposition()
erasure_result = self.subject_axis_erasure()
consistency_result = self.within_subject_direction_consistency()
return {
'identity_trap_risk': variance_result['is_identity_trap'],
'subject_variance_ratio': variance_result['subject_ratio'],
'erasure_gain': erasure_result['gain'],
'direction_consistency': consistency_result,
'recommendation': self.generate_recommendation(variance_result, erasure_result)
}
身份陷阱诊断框架揭示了 EEG 基础模型的 shortcut learning 问题。通过 FMScope 五项诊断，可以区分真实生物标志物与被试身份特征，为 EEG 基础模型的可靠性评估提供了标准化审计方法。这一发现对 EEG 临床应用和基础研究质量提升具有重要意义。