内容来源：https://www.wired.com/story/ai-isnt-smarter-than-a-baby-yet/

内容总结：

婴儿式学习启发AI进化：新测试揭示现有模型“力不从心”

尽管当前尖端人工智能模型能写代码、解高数、辩论哲学，但它们消耗海量数据和惊人能量的“笨重”学习方式，远不及一个1岁婴儿的高效。婴儿仅需一两次观察或短暂互动就能识别新物体，这种与生俱来的“轻量化”学习能力，正成为AI研究者寻求突破的关键灵感。

为探索这一前沿领域，Meta、斯坦福大学、东京大学及法国巴黎高等师范学院的研究人员共同开发了一项名为 EgoBabyVLM 的全新测试。该测试要求视觉语言模型（VLM）仅通过分析由婴儿和幼童头部摄像头采集的约1000小时视频（真实记录婴儿视角的“混乱”画面），来理解世界。结果令人震惊：当面对这些未经精心筛选、充满杂乱真实生活场景的素材时，主流顶级AI模型表现“惨败”，这暗示人类婴儿大脑独特的设计结构，使其能从未经处理的少量信息中快速学习。

斯坦福大学认知科学家迈克尔·弗兰克指出，婴儿的学习并非仅依赖语言，更源于丰富的多模态和触觉体验。例如，父母会谈论已不在视线中的物体，通过眼神或手势指物，或讨论过去与未来的事件。参与项目的研究人员认为，从认知科学和神经科学中借鉴“更长时间注意力”、“解读社交线索”等机制，有望推动AI向更类人的高效学习模式演进。

此前，2023年推出的BabyLM挑战已证明，基于Transformer架构的AI在语法学习上能用相当于10岁儿童的数据量表现良好，但专家强调，在理解物理世界、社交动态和“常识”方面，单纯依靠模式识别远远不够。普林斯顿大学的布伦丹·莱克表示，尽管有研究显示基础模型能通过婴儿视角数据学会“球”等简单概念，但距离儿童2岁时的全面推理能力仍有巨大鸿沟。

这种借鉴“婴儿式”学习偏好的新思路已初现曙光。斯坦福团队今年测试的新模型，利用婴儿头部视频数据，在学习物体因果关系、视觉与时间关系上效率显著提升。研究者认为，若模型能像人类一样天生偏向快速学习物理规律与社会关系，或将在整体上成为更高效的学习者。EgoBabyVLM挑战的推出，正激励全球研究者探索全新的架构与算法，试图让AI“返老还童”。

中文翻译：

如果你认为一台运行在数千块尖端计算机芯片上的人工智能模型很聪明，那不妨让我向你介绍一下“一岁婴儿”这个概念。
没错，婴儿也许不会编写计算机程序、解高等数学题或辩论哲学观点。但与当今那些消耗海洋般海量训练数据、能源消耗堪比小国家的人工智能模型不同，婴儿能以惊人的效率学习理解世界。他们看过一两次就能识别新物体，通过转瞬即逝的观察和身体互动来学习。
在改进人工智能方面，婴儿——以及他们大脑的结构——可能蕴含着关键启示。打造一个更像婴儿的人工智能版本，不仅能降低前沿模型的成本和能耗，而且如果未来由人工智能驱动的机器人要以更自然的方式学习环境，这种思路也会很有价值。
为了探索这片大胆的新领域，Meta、斯坦福大学、东京大学以及法国巴黎高等师范学校的研究人员开发了一项新测试，该测试突显了婴儿的学习能力，并推动人工智能研究者设计出能与婴儿能力相匹配的算法。
EgoBabyVLM挑战赛衡量的是视觉语言模型——即同时从文本和图像中学习的模型——在婴儿视角下理解世界的能力。该模型需要摄入约一千小时从固定在婴幼儿头上的摄像头采集的视频后，才能描述这个世界。（是的，这是真的。）
结果发现，这些尖端模型在面对这种真实又混乱的镜头时表现糟糕，这或许表明婴儿大脑的独特设计使其能从极少的信息中如此高效地学习。
婴儿并非从精心整理的数据集中学习，而是从万花筒般的视角中学习：父母谈论已经看不见的物体，用目光或手势指示东西，或讨论过去或未来的事情，而非当下正在发生的事。斯坦福大学专攻语言学习的认知科学家迈克尔·弗兰克参与了EgoBabyVLM的开发，他说婴儿不仅从语言中学习，还从丰富的多模态和触觉体验中学习。
这项测试表明，在人工智能领域，“显然需要的不只是语言，”弗兰克说。

语言学习

EgoBabyVLM只是科学家利用人工智能探索人类智能的最新例子。2023年推出的一项名为BabyLM的挑战任务要求人工智能模型学习语言的语法，使用的数据量大约相当于一个10岁孩子接触到的——数千万个单词，而人工智能模型则要处理数万亿个。值得关注的是，事实证明，基于Transformer的人工智能模型——通过关注不同句子中词语之间的关系来处理语言——在这方面表现得相当不错，这一发现对诺姆·乔姆斯基关于语法可能先天存在于人脑中的观点提出了挑战。
最初开发BabyLM的苏黎世联邦理工学院语言学家瑞安·科特雷尔表示，在理解物理世界方面，情况有所不同。“人类互动不会有庞大的语料库——没有人类互动的互联网，”他说。
麻省理工学院的认知科学家约书亚·特南鲍姆指出，BabyLM表明模型并未掌握关于物理世界、社会动态或心智理论的“常识”。
“Transformer非常擅长在数据中发现模式，”特南鲍姆说，“但似乎纯模式学习系统无法像婴儿或孩子那样接收数据，并学会他们所会的一切。”
一个长期存在的问题是，进化是否找到了优化人类及其他动物某些学习技能的方法，还是简单的学习算法就能做到我们做的一切。“在认知科学和神经科学中，关于大脑有多少是进化‘内置’的，存在很多争议，”特南鲍姆说，“大脑极其复杂，有很多内置的结构和架构。”
2024年，研究人员表明，一个基本的视觉语言模型仅通过摄入来自单个婴儿头部记录的数据，就能学会简单的东西，比如什么是球。但这距离以复杂方式推理世界还有很长的路。“谜团在于，儿童即使在两岁时是如何获得他们所拥有的全部能力的，”参与该项目的普林斯顿大学认知科学家布伦丹·莱克说。
EgoBabyVLM论文的作者建议，借鉴认知科学和神经科学的不同思路，可以推动更接近人类学习算法的进展。这包括设计能够保持长时间注意力并能解读社会信号的模型。
斯坦福大学的弗兰克已经证明，新方法可以让我们更接近婴儿式的人工智能。今年早些时候，他和同事用同样的婴儿头部视频数据测试了一种新型模型，该模型擅长学习因果关系、视觉与时间关系——即物体如何随时间相互影响。他们发现，这种新模型能更有效地学习不同物体的动力学特性，这是物理推理的基础。
这是一个令人着迷的可能性：也许那些更偏向于快速学习物理和社会关系等知识的模型，总体上可能成为更高效的学习者。
“EgoBabyVLM是一项精彩的挑战，”莱克说，“我很期待看到研究人员会提出哪些新的架构、方法和要素。”
本文为威尔·奈特“AI实验室”新闻通讯的一期内容。在此处阅读往期通讯。

英文来源：

If you think an artificial intelligence model running on thousands of cutting-edge computer chips is smart, allow me to introduce you to the concept of a 1-year-old.
OK, so babies might not be able to write computer programs, solve advanced math problems, or debate philosophical ideas. But unlike today’s AI models, which consume an ocean’s worth of training data and as much energy as a small country, babies learn to make sense of the world with amazing efficiency. They identify new objects after seeing them once or twice, and they learn through fleeting observation and physical interaction.
When it comes to improving AI, babies—and the architecture of their brains—might hold crucial insights. Building a more baby-like version of AI could make frontier models less costly and less energy intensive, and it might also be valuable if AI-powered robots are to learn about their environments in a more natural way.
To explore this bold new frontier, researchers at Meta, Stanford University, the University of Tokyo, and France’s École Normale Supérieure developed a new test that highlights the learning skills of babies and pushes AI researchers to design algorithms that match them.
The EgoBabyVLM Challenge judges how well vision language models, or VLMs, which learn from both text and imagery, can make sense of the world as a baby sees it. It requires a model to describe the world after ingesting about a thousand hours of video collected from cameras strapped to the heads of infants and toddlers. (Yes, really.)
It turns out that the cutting-edge models fail miserably when fed this realistic and messy footage, which suggests there may be something different about the design of the baby brain that enables it to learn so rapidly from so little information.
Instead of curated datasets, babies learn from a kaleidoscopic view of things: parents talking about objects that are no longer visible, indicating things using their gaze or a gesture, or discussing events from the past or in the future rather than whatever’s happening right then. Babies learn not just from language but also from a rich multimodal and tactile experience, says Michael Frank, a cognitive scientist at Stanford University who specializes in language learning and was involved with EgoBabyVLM’s development.
The test shows that when it comes to AI, “it’s clear that there’s more [than just language] that’s needed,” Frank says.
Language Learning
EgoBabyVLM is just the latest example of how scientists are using AI to explore human intelligence. A challenge called BabyLM, introduced in 2023, tasked AI models with learning the syntax of language using about the same amount of data a 10-year-old takes in—tens of millions of words, compared to trillions for AI models. Remarkably, it turns out that transformer-based AI models—which process language by paying attention to the relationship between words across different sentences—can do this quite well, a finding that challenges Noam Chomsky’s ideas concerning how syntax may be hardwired into the human brain.
Ryan Cotterell, a linguist at ETH Zurich who first developed BabyLM, says the situation is different when it comes to understanding the physical world. “There isn't going to be a large corpus of human interactions—there's no internet of human interactions,” he says.
Joshua Tenenbaum, a cognitive scientist at the Massachusetts Institute of Technology, notes that BabyLM showed models do not acquire “common sense” about the physical world, social dynamics, or theory of mind.
“Transformers are very good at finding patterns in data,” says Tenenbaum. “But it does seem that just pure pattern learning systems are not able to take the kind of data that a baby or a child receives and learn all the things that they do.”
An enduring question is whether evolution found a way to optimize certain learning skills in humans and other animals, or if simple learning algorithms can do everything we do. “There is a lot of debate in cognitive science and neuroscience about how much is built into the brain evolutionarily,” Tenenbaum says. “The brain is incredibly complex, and there's a lot of built-in structure and architecture.”
In 2024, researchers showed that a basic VLM can learn simple things, like what a ball is, purely by consuming data recorded from the head of a single infant. But this is a ways away from reasoning about the world in sophisticated ways. “The mystery is how children get to the full capabilities that they have even at the age of 2,” says Brendan Lake, a cognitive scientist at Princeton University who was involved with the project.
The authors of the EgoBabyVLM paper suggest that borrowing different ideas from cognitive science and neuroscience could enable progress toward more humanlike learning algorithms. This includes designing models that can pay attention over longer periods and can interpret social cues.
Stanford’s Frank has already shown that novel approaches can get us closer to baby-like AI. Earlier this year, he and colleagues tested a new kind of model that’s adept at learning causality and visual and temporal relationships—or how objects affect one another over time—using the same baby-head video data. They found the new model was able to learn about the dynamics of different objects, a foundation for physical reasoning, much more effectively.
It’s a tantalizing possibility: Perhaps models that are biased to learn more rapidly about things like physics and social relationships could be more efficient learners overall.
“EgoBabyVLM is a wonderful challenge,” says Lake. “I'm excited to see what kinds of new architectures, approaches, and ingredients researchers come up with.”
This is an edition of Will Knight’s AI Lab newsletter. Read previous newsletters here.