脑电专注力训练器（概念）
EEG attention trainer (concept)
一个长期的副项目，目前在慢炖模式。
为什么
大多数面向消费者的神经反馈应用奖励的是脑电特征：alpha/theta 比、前额叶不对称、SMR 上调。Attention as relation, not state 里我的论点是：这是把工具当成了判据。
这个项目是我尝试构建一台装置——任务才是判据，脑电只是一个让回路更快的工具。
当前架构草图
┌─────────────┐ ┌──────────────┐ ┌──────────────┐
│ Muse 2 / │ raw │ Preproc + │ feat. │ Task engine │
│ OpenBCI ├──────►│ bandpower ├──────►│ (game/work) │
└─────────────┘ └──────────────┘ └──────┬───────┘
│ task-
│ grounded
│ score
┌────────────────────────────────────────────▼───────┐
│ Reward shaper: 60% task perf + 40% EEG align │
│ (alignment is checked against a per-user baseline,│
│ not a population-mean signature) │
└────────────────────────────────────────────────────┘我要对抗的事：每个商业产品都把 EEG 对齐 当作主要奖励。我希望它是次要奖励，并且要在做完个人基线后才使用。
当前状态
- 数据流 + 频段功率提取流水线：✅ Muse 2 上跑通
- 个人基线校准：✅ 跑通（休息 + N-back 任务）
- 任务引擎：🚧 自适应难度的 dual-N-back 原型
- 奖励整形：🚧 在阅读 Goodhart 抗性奖励设计的文献，特别是从 RLHF 借灵感
- 硬件：长期偏好 OpenBCI Cyton 以获得 16 通道余裕；但 Muse 2 用于验证概念已足够
开放问题
- Transient hypofrontality 是有用的仪表对象，还是混杂因子（参见 心流与机器）。
- 上线方式：研究工具包（自带任务）还是打磨过的消费品（一个任务、调到极致）。
- 隐私：脑电是生物特征数据。默认本地优先，无显式同意不上云。
接下来
一个 dual-N-back 任务上的小公开原型，做足仪表化以便奖励整形器的行为可被检视。装置跑起来后，大概会写一篇配套随笔，谈 Goodhart 感知的奖励设计。
仓库：暂未公开。准备好后会与本站仓库并排存放。
A long-running side project, currently in slow-cook mode.
Why
Most consumer neurofeedback apps reward EEG signatures: alpha/theta ratios, frontal asymmetry, SMR uptraining. The argument in Attention as relation, not state is that this is an instrument mistaken for a criterion.
This project is my attempt to build a rig where the task is the criterion and the EEG is merely an instrument that gets the loop faster.
Architecture (current sketch)
┌─────────────┐ ┌──────────────┐ ┌──────────────┐
│ Muse 2 / │ raw │ Preproc + │ feat. │ Task engine │
│ OpenBCI ├──────►│ bandpower ├──────►│ (game/work) │
└─────────────┘ └──────────────┘ └──────┬───────┘
│ task-
│ grounded
│ score
┌────────────────────────────────────────────▼───────┐
│ Reward shaper: 60% task perf + 40% EEG align │
│ (alignment is checked against a per-user baseline,│
│ not a population-mean signature) │
└────────────────────────────────────────────────────┘The thing I’m fighting against: every commercial product makes EEG-alignment the primary reward. I want it to be the secondary reward, and only after individual baselining.
Current status
- Streaming + bandpower extraction pipeline: ✅ working with Muse 2
- Per-user baseline calibration: ✅ working (resting + N-back tasks)
- Task engine: 🚧 prototype dual-N-back with adaptive difficulty
- Reward shaper: 🚧 reading literature on Goodhart-resistant reward designs, especially scrutinising RLHF for ideas
- Hardware: prefer OpenBCI Cyton long-term for 16-channel headroom, but Muse 2 is fine for proving the concept
Open questions
- Whether transient hypofrontality is a useful instrumentation target or a confound (see 心流与机器).
- Whether to ship as a research toolkit (BYO task) or a polished consumer artefact (one task, well-tuned).
- Privacy: EEG is biometric data. Local-first by default, no cloud sync without explicit opt-in.
What’s next
A small public prototype on the dual-N-back task, instrumented enough that the reward shaper’s behaviour can be inspected. Probably a companion essay on Goodhart-aware reward design once the rig is running.
Repo: not yet public. Will live alongside this site’s repo when ready.
💬展开评论 / Show comments