来源：《新科学家》
原文刊登日期：2021年5月12日
An artificial neural network can interpret signals from the brain of a person who is imagining that they are writing with a pen, and convert them into text. The device converts words accurately at 90 characters per minute, more than twice the previous record for typing with a head- or eye-tracking system.
人工神经网络可以解读想象自己在用笔写字的人的大脑信号，并将其转化为文本。该设备能以每分钟90个字符的速度准确转换单词，是之前用头部或眼球追踪系统打字记录的两倍多。
These trackers allow people to move a mouse cursor and slowly type messages, but Jaimie Henderson at Stanford University in California says they are all-consuming for the operator. “If you’re using eye tracking to work with a computer then your eyes are tied to whatever you’re doing,” he says. “You can’t look up or look around or do something else. Having that additional input channel could be really important.”
这些追踪器可以让人们移动鼠标光标，慢慢地输入信息，但加州斯坦福大学的杰米·亨德森说，这些追踪器对操作者来说都是消耗。他说：“如果你在电脑上使用眼动跟踪技术工作，那么你的眼睛就会和你正在做的任何事情联系在一起。你不能抬头或环顾四周或做其他事情。拥有额外的输入渠道非常重要。”
To solve this problem, he and his colleagues implanted two small arrays of sensors just under the surface of the brain of a 65-year-old man who has a spinal cord injury that left him paralysed below the neck since 2007. Each sensor array was able to detect signals from around 100 neurons – a fraction of the estimated 100 billion neurons in the human brain.
为了解决这个问题，他和他的同事在一名65岁男子的大脑表面下植入了两个小型传感器阵列。自2007年以来，该男子因脊髓损伤导致颈部以下瘫痪。每一个传感器阵列都能够检测到大约100个神经元发出的信号——这只是人类大脑中估计的1000亿个神经元的一小部分。
As the man imagined writing letters and words on a piece of paper, the signals were fed to an artificial neural network. Team member Krishna Shenoy, also at Stanford University, says that the sensors don’t target exact neurons because many thousands or millions may be involved in hand movement, but with the two arrays monitoring around 200 neurons there are enough clues within the data for the artificial neural network to build up a reliable interpreter of brain signals.
当这个人想象在一张纸上写字母和单词时，信号被输入到一个人工神经网络。同样来自斯坦福大学的团队成员克里希纳·谢诺伊说，传感器不会瞄准确切的神经元，因为成千上万的神经元可能与手部运动有关，但有了这两个阵列监测大约200个神经元，数据中就有了足够的线索，可以让人工神经网络建立一个可靠的大脑信号解读器。
Often a neural network is trained with several thousand pieces of example data, which in this case would be a recording of a brain signal while writing a certain letter. That works fine when large data sets already exist or are provided by automated systems, but in this case generating an archive that large wasn’t practical because the man would have had to think about writing thousands of letters. Instead, the team took examples of signals from the man’s brain while writing certain letters and generated additional copies with random noise added to build a synthetic data set.
通常，神经网络是用几千个样本数据来训练的，在这种情况下，这些数据是大脑在写某个字母时的信号记录。当大数据集已经存在或由自动化系统提供时，这种方法可以很好地工作，但在这种情况下，生成这么大的数据集是不现实的，因为这个人需要考虑写数千个字母。取而代之的是，在书写某些字母时，研究小组从该男子的大脑中提取信号样本，并在这些信号中添加随机噪声，以构建一个合成数据集。
The model the team created won’t translate to another person because the neural network is trained only on data from one individual, with sensors placed in an unrepeatable location.
该团队创建的模型不能用于另一个人，因为神经网络只对来自一个人的数据进行训练，传感器被放置在一个不可重复的位置。
Using this system, the man was able to type at 90 characters per minute, approaching the average of people his age when using a smartphone, which is 115 characters per minute. The output had a 94.1 per cent accuracy, which increased to more than 99 per cent when an autocorrect tool was used.
使用这个系统，这名男子每分钟可以输入90个字符，接近他的同龄人使用智能手机时的平均每分钟115个字符。输出的准确率为94.1%，如果使用自动更正工具，准确率将升至99%以上。
Previous brain-computer interfaces have been able to interpret large signals, such as those for arm movements, but until now haven’t been able to pick up on those for fine movements like handwriting.
之前的脑-机接口已经能够解释大的信号，比如手臂运动的信号，但直到现在还不能识别精细的的动作，比如手写。
The team hopes to build on the work to create a speech decoder for use by someone who can no longer speak but is likely to still have the neural pathways to do so.
该团队希望在这项工作的基础上，创造一个语音解码器，供那些不再能说话但可能仍有神经通路的人使用。