神经科学与心理学2020nature.com编辑精选
用编码器-解码器框架将皮层活动机器翻译为文本
Machine translation of cortical activity to text with encoder–decoder framework
进阶 · 约 45 分钟读完前置知识：seq2seq/编码器-解码器框架、ECoG 与 EEG 的区别、词错误率(WER)概念
导读
这篇论文用编码器-解码器（seq2seq）框架将大脑皮层电活动（ECoG）信号“机器翻译”成文本，报告词错误率低至约 3%。但实验基于侵入式皮层电极，且只在 30-50 个固定句子、约 250 词的受限词表内解码，距离通用语音/思维转文字还很远。
HN 关注点：脑机接口进展，同时讨论焦点集中在该结果的实际局限：侵入式电极、极小的句子集合和低数据量导致的过拟合，使其与真正的“读心”相去甚远。
HN 评论精选
这几条评论里：质疑 2 补充 1 亲历 1
补充指出 ECoG 侵入性太强难以商用，EEG 因信号衰减和低通效应无法达到同等效果
很高兴看到进展，但这离公众可用还有几十年。解码脑信号的最大难题之一是获取大量能检测局部脑区电压的传感器。本研究用的是 ECoG（皮层脑电图），需要把电极直接植入大脑表面；而几乎所有消费级设备用的是 EEG（头皮脑电）。商用 ECoG 几乎不可能，因为需要极具侵入性的开颅手术，出于伦理原因研究中的植入物多半是为了诊断已有危重疾病而放置的。EEG 解码效果差得多：传感器与大脑距离远导致信号不局部化；皮肤和颅骨是很好的低通滤波器，会滤掉 100Hz-2kHz 的有趣高频信号；且这些信号功率很低。ECoG 因为电极 literally 贴在大脑表面，才能可靠捕捉这些信号。
bognition查看英文原文
Really cool to see progress made here but this won't be available for public use any time soon (likely decades). One of the biggest challenges with decoding brain signals is getting a large number sensors that detect voltages from a very localized region of the brain. This study was done with ECoG (Eletro-Cortico-Gram) which involves implanting small electrodes directly on the surface of the brain. Nearly all consumer devices use EEG (Electro Encephelo Gram) which involves putting sensors on the surface of the skin. Commercially available ECoG is highly unlikely as it requires an extremely invasive brain surgery. For ethical reasons the implants from the study we're likely implanted to help diagnose existing life threatening medical issues. Decoding speech from EEG won't work as well as ECoG a number of reasons. First the physical distance between the sensors and the brain Means the signals you pick up aren’t localized. Second the skin and skull are great low pass filters and filter out really interesting signals at higher frequencies, 100-2KHZ. Additionally these signals have a really low signal power because they're correlated with neuronal spiking. ECoG does a really good job picking on these signals because the sensor is literally on the surface of the brain. Its really hard to pick up these signals reliably with EEG.
质疑质疑 3% WER 的含金量：本质是 50 句分类任务，数据量瓶颈使其远非通用解码
很酷，但要指出局限：专业语音转写词错误率约 5%，他们做到 3%，看似更好，但语音转写要处理无限多的句子（包括从未说过的），而该方法只需区分 30-50 个句子，且同一句子在训练集至少出现两次、测试集出现一次——逐词解码本质上只是绕了个弯的 50 分类问题。由于是侵入式采集、数据成本高，他们处于极低数据量场景。理论上如果句子数据集是 TB 级而非 KB 级，可能达到谷歌翻译级别的准确率，但这样的数据量近期不可能收集到，需要机器学习在数据效率上的巨大飞跃。他们探索了迁移学习，值得肯定；跨被试建模几乎肯定是实现精度飞跃的必要条件。
lars查看英文原文
This is cool. For those who are not super familiar with language processing, I think it's good to point out the limitations of what's been done here though. They mention that professional speech transcription has word error rate around 5%, and that their method gets a WER of 3%. Sure, but the big distinction is that speech transcription must operate on an infinite number of sentences, even sentences that have never been said before. This method only has to distinguish between 30-50 sentences, and the same sentences must exist at least twice in the training set and once in the test set. Decoding word-by-word is really a roundabout way of doing a 50-way classification here. It's an invasive technique, so they need electrodes on a human cortex. This means data collection is costly, so their operating in very low data regime compared to most other seq2seq applications. It seems theoretically possible that this could operate on Google translate level accuracy if the sentence dataset was terrabyte sized rather than kilobyte sized. That dataset size seems very unlikely to be collected any time soon, so we'll need massive leaps in data efficiency in machine learning for something like this to reach that level. They explore transfer learning for this, which is nice to see. Subject-independent modelling is almost certainly a requirement to achieve significant leaps in accuracy for methods like this.
亲历亲历者分享用消费级 EEG 头环做简单词语解码仅达 40% 准确率的经验
我 5 年前用 choosemuse.com 的冥想头环做过类似尝试，它便宜且数据接口可 hack。我写了个手机 app：从"hello world my name is hrgiger"里随机挑一个词显示，绿灯亮起时我想这个词并眨眼，头环能检测眨眼，app 就用眨眼时间戳生成训练数据。我用这种方式做了几千条 6 分类训练数据，用自己半吊子的神经网络训练后预测，准确率从未超过 40%，试过混合脑波、原始数据、不同时间窗。仍是个有趣的项目。如果他们做出实用方案，我可以用整部影片字幕做全量训练，一个 Netflix 浏览器插件也许就行，但我不确定会不会有单一算法能理解每个人不同的数据。
hrgiger查看英文原文
I have tried something similar 5 years ago using the meditation device from choosemuse.com. It was the cheapest option and provided hackable interface that you had access all the data. Then I wrote a small mobile app connects to headset. Application was picking and showing a single random word from "hello world my name is hrgiger" then showing a greeen light, when I see green light, i think about the word and blink, headset was able to detect blinks as well so app was creating training data using blink time - xxx millis. So I created few thousands training data with 6 class using this and trained with my half-ass nn implementation and used generated weights to predict same way via mobile. Never achieved higher than 40%, tried all mixed waves, raw data, different windows of time series. Yet still it was a fun project to mess with, still I try to tune this nn implementation. If they achieve a practical solution I would use subtitles for the full length training. Simple netflix browser plugin might do the trick,but I am not sure if there will be a single AI algo that would understand everyones different data.
质疑质疑朗读训练能否迁移到真实思维解码，认为眼动追踪等替代技术短期内更实用
看起来很酷，但他们的模型是用朗读印刷句子的人训练的——这真的能推广到把抽象思想转化为文字的过程吗？研究者自己也承认因训练数据匮乏模型容易过拟合，且只用了 250 词的词表。Neuralink 也有夸大成果的商业动机，所以我不太信。很高兴看到进展，但在相当长时间内，眼动追踪和 P300 拼写器这类技术可能更可靠、侵入性更小。
leggomylibro查看英文原文
It looks cool, but they trained their models on people reading printed sentences out loud. Would that actually translate to decoding the process of turning abstract thoughts into words? The researchers also note that their models are vulnerable to over-fitting because of the paucity of training data, and they only used a 250-word vocabulary. Neuralink also has a strong commercial incentive to inflate the results, so I'm not too sure about this. It's great to see progress in these areas, but it seems that technologies like eye-tracking and P300 spellers are probably going to be more reliable and less invasive for quite some time.
相关资源
从这篇论文所有 HN 讨论帖的全部评论里整理出来的链接。
历次讨论
- 2020Hacker News 讨论帖195 分 · 91 评
同领域还值得读
神经科学家能理解微处理器吗？（2017）
Could a Neuroscientist Understand a Microprocessor? (2017)
提出一个思想实验：把神经科学常用的分析方法（损伤实验、电生理、功能定位等）应用到一个我们完全了解其原理的工程系统——6502 微处理器（运行游戏）上，检验这些方法能否真正“理解”…
神经科学与心理学2017 · 411 分 · 135 评进阶 · 约 60 分钟编辑精选大脑衰老呈现非线性转变，提示中年“关键窗口”
Brain aging shows nonlinear transitions, suggesting a midlife "critical window"
这篇 PNAS 论文研究大脑衰老的非线性变化，发现大脑老化并非匀速进行，而是在中年（约 40-60 岁）存在一个关键的"临界窗口期"，此时干预可能有效。
神经科学与心理学2025 · 295 分 · 159 评进阶 · 约 75 分钟编辑精选社交网络中大规模情绪传染的证据（2014）
Evidence of massive-scale emotional contagion through social networks (2014)
这篇 2014 年发表于 PNAS 的论文研究了 Facebook 上约 70 万用户的信息流中，减少正面或负面内容是否会传染性地影响用户自己发布内容的情绪。
神经科学与心理学2020 · 285 分 · 72 评入门 · 约 45 分钟编辑精选裸盖菇素使人类大脑去同步化
Psilocybin desynchronizes the human brain
这篇 Nature 研究用个体化的 fMRI 精密成像追踪受试者服用裸盖菇素（psilocybin）前后的脑功能连接变化，发现药物会造成大范围的脑网络「去同步化」，尤其是默认模式…
神经科学与心理学2024 · 273 分 · 259 评进阶 · 约 60 分钟编辑精选对智力的赞扬会削弱儿童的动机与表现（1998）
Praise for Intelligence Can Undermine Children's Motivation (1998)
这篇 1998 年的经典论文（Dweck 等人）通过实验研究表扬方式对儿童动机的影响：夸孩子“聪明”（表扬智力）会让他们回避挑战、害怕失败，而表扬努力则更能培养坚持性。
神经科学与心理学2016 · 253 分 · 155 评入门 · 约 45 分钟编辑精选