人工智能：拥有不同于人类思维的机器

人工智能：拥有不同于人类思维的机器

Artificial intelligence: The machines with alien minds

Our smartest machines look nothing like we predicted – has the field lost its way, or do we need to rethink what AI actually means, asks Tom Chatfield.
Would modern artificial intelligence live up to the dreams of the field’s founders? Perhaps not. But in many ways, the smartest machines we have built are entities they never could have imagined.

In 1956, attendees of a research camp at Dartmouth College in New Hampshire coined the phrase "artificial intelligence" to describe its efforts to “find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.”

Compare that with the AI project that Facebook announced this month. Under one of the world’s most prominent experts, it will “do world-class artificial-intelligence research using all of the knowledge that people have shared on Facebook” – with potential gains including building “services that are much more natural to interact with.”

What does that mean? Like much of modern AI, they will be training algorithms to sift and analyse unimaginably vast amounts of data in the hope that smart answers will emerge. Google, IBM and many others are now using this technique – called machine learning – to great commercial success, and the “intelligence” they create underpins everything from your internet searches to online language translation. Since the calculations of these machines involves making statistical correlations within huge caches of data, their reasoning can be unfathomable to the human mind – often these systems provide apparently intelligent answers, but nobody has any idea how they came to their conclusions.

Yet even the most advanced forms of machine intelligence cannot hope to pass for a human in Turing’s famous test – let alone use natural language or develop concepts themselves, as the pioneers hoped. More than half a century of research has brought us a far more sophisticated grasp of what machine intelligence looks like – but has it lost its way? Or do we need to reframe our ideas about what the term AI actually means?

One person who believes progress in AI has fallen short in many ways is the author and academic Douglas Hoftstadter – most famous for his Pulitzer-Prize-winning 1979 book Gödel, Escher, Bach – who in a recentprofile for The Atlantic magazine emphasized his disillusionment with the current direction of AI.

For Hoftstadter, the label “intelligence” is simply inappropriate for describing insights drawn by brute computing power from massive data sets – because, from his perspective, the fact that results appear smart is irrelevant if the process underlying them bears no resemblance to intelligent thought. As he put it to interviewer James Somers, “I don’t want to be involved in passing off some fancy program’s behaviour for intelligence when I know that it has nothing to do with intelligence. And I don’t know why more people aren’t that way.”

Cheap tricks

By Hofstadter’s standards, iconic computational achievements like beating the world’s best players at chess or Jeopardy are rendered trivial by the “trickery” involved: by the fact that the winning computer has done little more than weigh the relative benefits of several billion possibilities, without at any point knowing anything about the nature of the game being played.

What computers scientists should be researching, he argues, is how the phenomenon of intelligence itself arises from the human brain – a question that those focused on data-led machine analysis (and its profitable applications) have increasingly side-lined.

How far current AI research can take us is open for debate, but even its proponents acknowledge that the current emphasis on data and pragmatism may only take us so far. In an analogy quoted by The Atlantic from a leading textbook in the field, it’s like trying to climb a tree to the moon. Progress is excellent at first – but cannot continue past a certain point. Years of success may remain in feeding vast quantities of problems and solutions into machines running clever learning programs, and watching them train themselves to produce good-enough outcomes. But what next, when that particular tree runs out? And where is it we should be aiming for in the first place?

Despite Hoftstadter’s protestations, perhaps we need to stop comparing machines with ourselves altogether. After all, while humans are the only examples we possess of phenomena like advanced language and logic, they’re hardly the earth’s only examples of intelligence. From pet parrots to ant colonies, our world is packed with “intelligent” responses.

How do we do that? A first step may be to change the language we use to describe these machine minds. Today, the very phrase “artificial intelligence” conjures certain iconic images; a robotic reflection of ourselves. Yet this hypothetical being stands amid an ocean of vagueness. From the phenomenon of consciousness to what it might mean to measure or analyse intelligence, our own minds remain profoundly mysterious – and those insights we do have are rooted in millennia of evolutionary history, the intricacies of our synapse-packed brains, and investigations spanning every field from literature to the social sciences.

Why should our biological manner of thinking determine our approach to silicon-based circuits and electronic logic? Our machine creations are more profoundly divided from us than anything else in nature. They do not need to think like us to serve us, work with us, or even understand us – as our own relationships with nature should teach us at a glance.

So what new words might we use in place of artificial intelligence? “Artificial” implies something bogus, ersatz and somehow unreal – which is why I feel the term “machine” fits better. “Intelligence” implies discernment and apprehension, but also something inexorably human – which is why I prefer the more impassive “reasoning.” Machine reasoning: it’s not a phrase for the ages, but perhaps a beginning to the process of seeing what waits under our noses.

我们所发明的最聪明的机器与我们所预测的截然不同，难道这一领域失去方向了，亦或者我们需要重新思考人工智能的真正含义？

现代人工智能是否会达成该领域创立者的梦想？答案也许是否定的。但是我们所建造的最聪明的机器已经在很多方面完全超越了他们的想象。

1956年，新罕布什尔州达特茅斯学院某研讨营的参与者创造了“人工智能”这个短语，用来形容某一活动，它旨在“制造出一种机器，这种机器能使用语言，形成抽象概念和观念，拥有人类所独有的解决问题的能力并可以完善自己。”

再看看Facebook于3月宣布的人工智能计划。在世界顶尖级专家的参与下，该计划会“利用人们在Facebook上分享的知识进行世界级的人工智能研究”，潜在收益包括打造出“互动起来更自然的服务”。

这意味着什么？像大多数现代人工智能研究那样，专家们会让计算程序不断筛选和分析难以想象的大量数据，以期能得到聪明的答案。谷歌、国家商用机器公司等正利用机器学习这项技术，并获得了巨大商业成功，而它们所创造的“智能”维系了从网络搜索到在线语言翻译的一切网络服务。这些机器在计算时需要对大量数据作出统计相关性分析，所以它们的推理过程难以为人类所了解，通常情况下，机器会给出明显很智能的答案，但是没有人清楚它们是怎么得出结论的。

但是即使是最先进的机器智能也无法通过图灵的著名测验，够不上拥有人类的思维，它们更做不到人工智能创始者所希望的那样：使用自然语言或是自己发展出概念思维。半个多世纪的研究让我们更深入地了解了机器智能的大概，但是它是否已经迷失方向？或者我们是否需要重新思考人工智能这个词语的真正含义？

有人认为人工智能的研究在很多方面低于预期水平，作家兼学者道格拉斯·霍夫斯塔特（Douglas Hoftstadter）就是其中一位，他撰写的书籍Gödel, Escher, Bash获得了1979年普利策奖。在《大西洋月刊》对他的最新介绍中，他强调对人工智能当前的发展方向失望。

对霍夫斯塔特而言，“智能”这个标签已经不适用于描述通过毫无理性的计算能力从数据集得出的结论，因为在他看来，如果得出结论的过程与智能思维截然不同，即使结论看上去很聪明，也与智能挂不上钩。詹姆斯·萨默斯（James Sommers）采访霍夫斯塔特时，他是这样说的：“我不想把一些高级程序的行为当成是智能，因为我知道那压根和智能扯不上关系。我不明白为什么越来越多的人不是这样认为。”

廉价的把戏

按照霍夫斯塔特的标准，计算机取得的标志性成就都微不足道，例如在围棋或“挑战自我”游戏上打败世界最厉害的选手，因为其中的“戏法”不过是：获胜的电脑除了衡量几十亿种可能性的相对好处外，什么事都没做，在整个游戏过程中完全没有了解所玩游戏的性质。

霍夫斯塔特争论道，计算机科学家应该研究人脑是如何产生智能的，而这个问题越来越受到那些注重运用数据进行机器分析（并研发效益高的应用）的研究者的忽视。

当前的人工智能研究能走多远尚待讨论，但是即使是它的支持者也承认目前对数据和实用主义的强调会让研究走不长远。《大西洋月刊》从该领域的顶尖教材中引用的类比称，这样的研究就像试图爬树去月亮。一开始研究取得了极佳的进步，但是到达某点后就无法持续下去。多年的成功也许还是停留在向机器输入大量问题和解决方案，一旁看着，希望通过运行聪明的学习程序，这些机器能自我训练，得出好结果。但是如果这棵树到头了该怎么办？我们一开始便应该瞄准的目的地是哪？

霍夫斯塔特说的先不管，也许我们应该停止将机器和我们自身相比。毕竟，尽管我们认为人类是唯一拥有高级语言和逻辑的实例，他们不会是地球上唯一拥有智能的生物。从宠物鹦鹉到蚂蚁王国，我们的世界充满了“智能”的反应。

我们要怎样做到呢？首先也许是改变我们用来描述这些机器头脑的语言。现在“人工智能”这个短语会唤起特定的标志性形象：人类的机器版本。然而，这个假定的存在还是处于一片混沌中。我们既弄不清楚意识现象，也不懂衡量和分析智能意味着什么，我们的见识都植根于几千年的进化史、突触密集的精细大脑和横跨文学至社会科学各个领域的研究。

为什么我们生物的思维方式要决定我们对待硅基电路和电子逻辑的方法？同其他事物相比，我们创造的机器本质上与我们更为不同。要服务我们，与我们共事，甚至理解我们，机器完全不需要像我们一样思考，我们与自然的关系立马就可以说明这一点。

那么，我们可以用什么新词取代人工智能？“人工”意味着虚假仿制，透露出不真实的感觉，因此我觉得“机器”这个词语更恰当。“智能”意味着洞察力和理解力，与人类密不可分，因此我更喜欢用不惨杂感情的“推理”一词。机器推理并不是可以一直用下去的词语，但是却可能是另一个过程的开端，让我们看清近在眼前等待的事物。

>>点击查看更多关于双语新闻的文章