经济学人 | 科学家通过婴儿的视角训练AI

Scientists have trained an AI through the eyes of a baby

科学家通过婴儿的视角训练 AI

“Chair” and “ball” were among little AI’s first words

“椅子”和“球”是这个宝宝 AI 学会的首批词汇之一

For decades linguists have argued over how children learn language. Some think that babies are born as “blank slates” who pick up language simply from experience—hearing, seeing and playing with the world. Others argue that experience is not enough and that babies’ brains must be hardwired to make acquiring language easy.

几十年来，语言学家对儿童是如何习得语言的争论不休。一些人认为婴儿生来“白纸一张”，仅凭体验(听、看和与外界互动)来学习语言。其他人则认为体验还不够，婴儿的大脑必然天生具备一些条件，能让语言习得变得容易。

AI models such as GPT-4 have done little to settle the debate. The way these models learn language—by trawling through reams of text data from millions of web pages—is vastly different to the experiences of babbling babies.

GPT-4 等 AI 模型在解决这个争论方面作用甚微。这些模型通过搜罗千百万网页上的海量文本数据来学习语言，这与婴儿咿呀学语的过程截然不同。

A team of scientists at New York University examined the question by training an AI model on the experiences of a single infant . Between the ages of six and 25 months, a toddler called Sam wore a head-mounted camera for an hour a week—around 1% of his waking hours. The camera recorded everything he saw and heard while he played with toys, enjoyed days at the park and interacted with his pet cats.

纽约大学的一个科学家团队对这个问题做了研究，他们根据一个婴儿的学语过程来训练AI模型。在一个名叫山姆的幼儿从6个月至25个月大的这段时间里，科学家让他佩戴头戴式摄像机，记录下他玩玩具、去公园以及与家中宠物猫互动时的情景，每周共戴一小时(约占他醒着的时间的1%)。

The recordings and transcribed audio were fed into an AI, which was set up to know that images and words that appeared at the same time were related, but was otherwise left to make sense of the mess of colours and speech that Sam experienced.

摄像机记录下来的场景和根据音频转写的文字被输入到一个 AI 模型中，AI 根据设定知道同时出现的图像和单词有关联，但它要自己理解山姆看到的各色图案和听到的话语。

Despite the limited training data, the AI was able to pick out objects and learn the matching words. The researchers tested the model by asking it to identify objects that Sam had seen before, such as a chair from his home or one of his toy balls. Given a list of four options the model picked the correct word 62% of the time, far above the chance level of 25%.

尽管训练数据有限，AI 模型仍能够辨别出物体，并学习对应的单词。研究人员要求模型识别山姆曾看到过的物体，以此测试它的学习效果，比如要求它识别山姆家里的一把椅子或者他的一个玩具球。面对四个备选词，模型的选择准确率达到了62%，远高于25%的随机水平。

To the researchers’ surprise, the model could also identify chairs and balls that Sam had never seen. The AI learnt at least 40 different words, but it was far from matching Sam’s vocabulary and language abilities by the end of the experiment.

让研究人员惊讶的是，模型还能识别山姆从未见过的椅子和球。AI模型至少学会了40个单词，但在实验结束时，它的词汇和语言能力与山姆相去甚远。

The researchers, published recently in the journal Science, argue that, to match words to objects, learning from experience may well be enough. Sceptics, however, doubt that the AI would be able to learn abstract nouns or verbs, and question how similar the learning processes really are. The mystery of language acquisition lives on.

这项研究最近发表在《科学》杂志上，研究人员认为，要将单词与物体匹配起来，仅凭体验学习可能已经足够。然而，持怀疑态度的人认为 AI 无法学习抽象名词或动词，并质疑 AI 这一学习过程与幼儿学语真有多少相似性。语言习得的谜团仍然未解。