科研 Skill

神经信号基础模型评估

neuro-foundation-model-eval

Fairly evaluates pretrained EEG or iEEG foundation models against classical baselines with matched data and splits. Use when fine-tuning or linear-probing a pretrained model.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Foundation-model evaluation for neural signals

Pretrained EEG models are often compared against weak baselines, on splits that leak, or on data seen during pretraining. A fair comparison holds everything fixed except the model.

1. Check for pretraining overlap

List the pretraining corpora of the model (from its paper or card). If the downstream dataset, or recordings of the same people, are in the pretraining set, results are contaminated. Prefer datasets clearly outside pretraining, or say plainly that overlap exists.

2. Same data, same split, same budget

  • Use subject-wise splits (see neuro-bci-decoding-eval) identical for every method, ideally published folds.
  • Match preprocessing to each model's expected input (sampling rate, channel set, units, window length) and document resampling or channel mapping, since mismatches silently hurt the pretrained model.
  • Give baselines the same tuning effort and the same labelled data amount.

3. Baselines that must appear

  1. Classical pipeline suited to the task: band power + LDA/logistic, or Riemannian tangent space (pyRiemann) for motor imagery/ERP.
  2. A compact supervised CNN trained from scratch (EEGNet or ShallowConvNet via braindecode).
  3. The pretrained model, linear probe.
  4. The pretrained model, full fine-tune.

4. Measure what pretraining should buy

  • Label efficiency: performance vs number of labelled subjects or trials (for example 1, 5, 10, 25, 100%). Pretraining should help most at low labels.
  • Transfer: new dataset or device not seen in pretraining.
  • Cost: parameters, GPU time to fine-tune, inference latency.

Run 3 or more seeds and report mean and spread; differences smaller than seed variation are not gains.

Report

Pretraining corpora and overlap check, downstream data and split, input harmonisation, all baselines with tuning budget, metric with spread across seeds and subjects, label-efficiency curve, compute cost.

示例提示词

帮我公平评估这个预训练EEG基础模型相比传统方法有没有真的提升。

所属插件

神经科学面向神经科学研究:EEG/MEG、fMRI、BIDS、spike sorting、连接组、脑机接口解码评估、神经刺激安全与公开神经数据检索。

安装该插件会一并获得同领域的 45 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

神经信号基础模型评估能做什么?

Fairly evaluates pretrained EEG or iEEG foundation models against classical baselines with matched data and splits. Use when fine-tuning or linear-probing a pretrained model。

如何安装神经信号基础模型评估?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「神经信号基础模型评估」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

神经信号基础模型评估需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

神经信号基础模型评估经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。