需用时 06:19

果壳网:你使用谷歌翻译(Google Translate)阅读果壳网对道金斯的访谈文章。谷歌翻译是与你相对的研究思路的象征——大数据、产业AI、基本不涉及“思考”。你怎么看谷歌翻译这个产品?它所代表的研究趋势有可取之处吗?你认为要解开人类思考运作机制的原理,关键在哪里?


Guokr: You used Google Translate to read our interview with Dr. Dawkins.  Google Translate seems like a typical example that uses the opposite idea of your research — all big data, all industrial AI, hardly any “thinking”.  What do you think of Google Translate?  Do you think there is an upper limit to the trend it represents?  If yes, where could it lie?  Will your research helps to break the limit, and why?




Actually, I didn’t use Google Translate to read the Dawkins interview — I read it in Chinese myself.  I’ve studied Mandarin for a long time, and although I am by no means fluentin it, I can read some things, albeit slowly.

You are quite correct, however, in saying that Google Translate represents a philosophy that is diametrically opposed to my own.  Several decades ago, I was profoundly drawn to the very young field of artificial intelligence because I was fascinated by the question “What is thinking?”  When I started devising my own AI research projects, back in the late 1970s and early 1980s, I was totally driven by this question.  As a professor, I also taught a course on AI, and although at first I liked the textbooks I used, after a couple of years I was totally disillusioned.  They seemed to contain empty recipes for achieving impressive goals in ways that bypassed thinking rather than imitating it.  Thus, for instance, following the textbooks, I taught my students chess-playing algorithms based on a brute-force mathematical technique called “minimax lookahead with alpha-beta pruning”, and although such techniques were ingenious and quite fun to teach, and although they certainly led to excellent chess-playing programs, they clearly had nothing in common with what goes on in the mind of a grandmaster (or even in the mind of a human beginner).  Because of realizations like this, I lost all interest in standard AI within a couple of years, and my own research became entirely based on techniques that had practically nothing at all to do with the ideas I had found in AI textbooks and had taught to my students.

从那时起,我在认知科学方面的研究一直是关于模仿微领域下的人类思维,譬如涉及字母串的“有样学样”领域(Copycat domain)。(在这里应当把字母看成26个从az拥有天然顺序的抽象体,而不是具体的形状或声音。)我相信思考牵涉到找出情境的本质,这一理念与类比是认知的核心这一观点密切相关。基本上讲,这一观点认为当我们忽略表象,找到了新旧两种情境共有的深层核心本质,而将新的情境视为与之前的某个情境(或某一类情境)“一样”时,就产生了类比。因此,微领域下的类比就是我过去30年一直在研究的东西。多年来,我想出了好多种不同的微领域,每一种都自有其迷人之处,但为了简洁,接下来我单就“有样学样”领域谈一谈。


Ever since then, my research in cognitive science has always been about imitating human thinking in microdomains, such as the Copycat domain, which involves strings of letters.  (One should think of the letters as 26 abstract entities that have a natural order, running from a to z, rather than as physical shapes or sounds.)  My belief that thinking involves finding the essence of situations is closely related to the idea that analogy is the core of cognition.  The basic idea is that an analogy happens when we perceive a new situation to be “the same as” some previous situation (or class of situations) by ignoring superficial features and finding a deep central essence that the new and old situations share.  Thus the study of analogies in microdomains is what my research has been about, for the past three decades.  Over the years I have devised quite a few diverse microdomains, each of which has its own fascination, but for the sake of brevity, I’ll focus only on the Copycat domain.


假如 abc 变为 abd,那么 icjjckkkc 该变成什么?

实际上说它是“类比问题”容易引起误会,因为“问题”这个词暗含了它有唯一正确解的意思——但这个推论完全是错误的!其实该将上面这个句子称为“类比疑问”或“类比疑难”,而且需要意识到它有很多不同的合理 答案,这些答案没有对错之说。好了,现在我再次希望每一位读者在读下去之前都尽力去想一个或更多的答案。所以,现在请停止阅读!


Here is a typical analogy problem in the Copycat microdomain for Guokr’s readers to think about:                        If abc changes to abd, then what should icjjckkkc change to?

Actually, calling this an “analogy problem” is misleading, since the term “problem” tends to suggest that it has a single correct solution — but that conclusion is totally wrong!  In fact, one should call the above an “analogy question” or an “analogy challenge” instead, and one should recognize that it has a wide range of plausible answers, none of which is correct or incorrect.  And so now, I once again encourage each of your readers to try to think of one or more answers to this challenge before going on.  Therefore, please stop reading now!

好——我们接着说。我希望你至少想到了一个答案,要是想到了两个、三个甚至更多就更好了!下面是我就这个学样疑难想到的几个可能的答案: icjjckkkd、 idjjdkkkd、 icjjckkdc、 icjjcdddc、 icjjclllc、 icjjckkkkc  和 icjjcllllc。 (我还能列出更多,但有这些就足够了。)仔细检查你会发现,每一个答案都有它自己的“内在逻辑”自圆其说(例如,第一个答案将最右边的 变成了d,可以说正好是 abc 变成 abd  时发生的情况)。不过,尽管每个答案都有“逻辑性”,但大多数人都还是会觉得有的答案弱一些、有的强一些或是令人信服,对他们而言,“最佳”答案拥有一种难以捉摸、不可名状的审美意义上的“魅力”。你最喜欢哪个答案?你有没有想出跟我列的这些都不同的答案?如果有的话,那可真有意思了!


All right — let’s go on.  I hope you found at least one possible answer on your own, and hopefully two, three, or more!  Here are a few of my own possible answers to this Copycat challenge: icjjckkkd;, idjjdkkkd, icjjckkdc, icjjcdddc, icjjclllc, icjjckkkkc, and icjjcllllc.  (I could list many more, but this will suffice for now.)  As you will see if you inspect them, each one has its own defensible brand of “internal logic” (the first one changes the rightmost c to a d, for instance, which is arguably exactly what happened when abc changed to abd), and yet, despite the “logicality” of each answer, most people tend to see certain answers as weak and others as strong or even compelling, with the “best” ones having an elusive, ineffable, esthetic “charm” to them.  Which answer do you prefer?  Did you come up with one or more answers different from all the ones in my list?  If so, that’s very interesting!

在上面所列的7个答案中,我最喜欢后面的3个,而且说实话,我也说不好哪一个我最喜欢。本能也好、直觉也罢,我认为 icjjckkkc 字符串里的3个“c”只是一种“不重要的黏合剂”,起到分隔3个“更加重要的”部分——i、jj 和 kkk——的作用。我无法证明这一点,但我就是这么觉得。换句话说,对我而言,9个字母的字符串 icjjckkkc “本质上”只是3个元素的字符串——i-jj-kkk——因此我需要转换的部分就成了最右边的元素,即 kkk(当然,我可以只改变最右边的 k,也就是上面第4个答案那样,但我认为这样改太丑了,所以我不这么做——即使我充分意识到这么做的逻辑!)然后问题就成了,“如何改变 kkk?”当然,一种方法是将其在字母顺序上往前进一位(模仿 c → d 的转变),因此就有了 kkk → lll,对应的答案是 icjjclllc。另一种看问题的思路关注3组元素 ijj 和 kkk 长度——也就是关注其中暗含的数字模式“1–2–3”。这种数字模式正正好与字母模式 abc 对应,因此,试着像转换 abc → abd  那样,可以想到将“1–2–3”变成“1–2–4”,而这样就得出了 icjjckkkkc 这个答案。但还有一种可能,就是将刚刚提出的两种答案加起来,于是乎不仅 变成了 l,3也变成了4,于是就得到了icjjcllllc

好笑的是,有的人可能会认为这最后一个解是“给百合花镀金”。他们会说,“把 变成 可以,把3变成4也行,但把两个加起来一次都变就说不过去了,这就跟在冰淇淋上洒辣椒粉或是画完蛇再添上足没什么两样。”


In the list of seven answers above, my favorites are the last three, and to tell the truth, I’m not sure which of them I prefer.  Instinctively or intuitively, I feel that the three c’s inside the string icjjckkkc are merely a kind of “unimportant mortar” serving to separate three “more important” items — namely, i, jj, and kkk.  I can’t prove this, but it’s just how I feel.  In other words, for me, the nine-letter string icjjckkkc is “essentially” just a three-element string — namely,  i-jj-kkk — and so the item I wish to change is the rightmost one, namely kkk.  (Of course, I could just change the rightmost k, as in answer #4 above, but to me that idea seems very ugly, so I don’t go for it — although I certainly recognize its logic!)  Then the question is, “How to change the kkk?”  One way, of course, would be to advance it alphabetically (mimicking the c→ d change), which would give kkk→ lll, which corresponds to the answer icjjclllc.  A rival viewpoint focuses on the increasing lengths of the groups i, jj, and kkk — that is, it focuses on the hidden number-pattern “1–2–3”.  Now this number-pattern corresponds very precisely and elegantly to the letter-pattern abc, and so, in trying to imitate the change abc→ abd, one is led to the idea of turning “1–2–3” into “1–2–4”, which would yield the answer icjjckkkkc.  But there is a third possibility, which involves combining the two answers just proposed, so that not only does k become l, but also 3 becomes 4, so that one would get icjjcllllc.

Amusingly enough, some people might call this last solution “gilding the lily”. They would say, “Letting k turn into l is fine, and letting 3 turn into 4 is also fine, but putting the two ideas together and doing both actions at once makes no more sense than combining chili peppers with ice cream, or adding feet to one’s drawing of a snake.”



Which answer do you prefer?  I’m not sure, myself, but what I can say for sure is that my ultimate choice, whatever it is, will be the winner in an “esthetic battle” taking place inside my head.  It is not a question of logic or truth, but a question of beauty, which means it all comes down to taste.  This way of thinking about the mind is a far cry from what most AI researchers have been thinking for the past 50 or 60 years, and it is unpopular among them because it is so unclear how to go about computationally modeling esthetic taste, whereas trying to use logic or mathematics to imitate thinking seems so much more straightforward.  However, those formal kinds of methodologies have led to extremely brittle kinds of “intelligence”, lacking all insight.  They are “barking up the wrong tree”, in my view.

我的研究小组叫做“灵活类比研究小组”(Fluid Analogies Research Group,FARG),这里的研究肆无忌惮地涉及了各种各样并排进行的相互竞争的微压力,其整体结构类似蚁丘。没错,我说的是计算模型,但与AI研究中通常做的那些大相迥异,也跟神经网络十分不同。FARG系统(比如 Copycat、Tabletop、Letter Spirit、Phaeaco、Seqsee等)的智能是集合智能,每一种都是从成千上万种小得不能再小、微得不能再微的活动中净胜而出,单独看它们并不是智能的。



The research done in my research group, called the “Fluid Analogies Research Group” (FARG), shamelessly involves all sorts of micro-pressures competing with one another in parallel, in a kind of anthill-like architecture.  Yes, I am speaking of computational models, but of a very different sort from what is standardly done in AI, and also very different from neural networks.  The intelligence of FARG systems (such as Copycat, Tabletop, Letter Spirit, Phaeaco, Seqsee, and others) is collective and comes out of the competition among thousands of extremely myopic and tiny activities that are not at all intelligent on their own.

And let me also stress that the word “intelligence” here has to be taken with a huge grain of salt.  Our computational models of cognition are far from being truly intelligent, even in the tiny microdomains in which they operate.  They often give hilariously silly-seeming answers to questions, because we haven’t managed to impart the proper mechanisms or pressures or tastes to our programs — and watching these failures of our programs delights us rather than saddening us, because it reminds us of the amazingly subtlety and elusiveness of the human mind.  We would be profoundlly sad if any of our systems actually had attained human-level intelligence in its microdomain, because that would be scary: it would suggest that human intelligence is not nearly as complex or deep as we had thought.  It would suggest that just a few decades of research had sufficed for humanity to unravel all the mysteries of the human mind, and that, to my way of thinking, would be tragic.



We in FARG are not trying to develop practical applications such as translation engines or question-answering machines or Web-search software or anything of the sort.  We are simply trying to understand the nature of human concepts and the fundamental mechanisms that underlie human thinking.  We are more like philosophers or psychologists trying to probe the secrets of the human mind, rather than engineers trying to make clever computers or smart programs.  We are old-fashioned purists who are driven by a deep philosophical curiosity, not by the desire to make practical devices (let alone the desire to make huge amounts of money!).

思维的类比:面对这么多不同的形状,人类还是能很容易就判断出这是字母“A”。在这种纷繁多样之下,有着怎样的抽象机制?侯世达,“On seeing A's and seeing As”,SEHR, volume 4, issue 2: Constructions of the Mind,Updated July 22, 1995

果壳网:丹尼尔·丹尼特(Daniel Dennett)曾说过,人脑和电脑的差异可能在于电子元件不像神经元那样“竞争”,二极管不会为了电力而求生。你觉得硬件上的差异重要吗?人工智能依赖于硬件这样的物理机制吗?


Guokr: Daniel Dennett once remarked that diodes don’t compete for resources like neurons do, so computer never face true struggle, and will never actually “want” anything.  Do you think such a hardware difference matters?  Does AI depend on certain underlying physical mechanisms?

侯世达:我没见过丹尼特的这句话(顺便说,丹尼特是我的好朋友、合著者——三十多年前丹和我一起编撰了《思维的自我》[The Mind’s I]这本书,我们合作非常愉快!),我不知道对此有何看法。我自然明白其背后的思想,但我不确定真正的思考是否真的依赖于细到神经元层面甚或更下层的处理过程。要说“思考确实有赖于所有那些非常细粒、非常底层的生物硬件”,这种观点本身当然是可以想象的;但在我看来,同样完全可以想象,思考依赖的是某个更高层的潜认知处理过程,而那些较低层的东西,比如神经元和神经递质,仅仅是思考可能的基质之一。丹尼特可能会认为,倘若有人想创造一台思考机器,神经硬件是不可或缺的——但我对此观点表示怀疑。但假如他真的这么说了,那我很好奇,如果我们造出了拥有数千亿神经元的极度精确的计算机网络,他会作何评价。在这个模型里不会有物理上的神经元,只有千亿复杂数据结构以微妙的方式相互触发,其行为方式和神经元完全一样,但物质组成却丝毫不同。我认为如果这个模型能通过图灵测试,那么丹尼特(和约翰·塞尔[John Searle]不同,后者因他的“中文屋”[Chinese room]思想实验而出名,但我认为这个思想实验从头到脚漏洞百出)肯定会非常高兴地说它是真正在思考,它有意识、它说出的话和人类说的话一样充满了意义。我也会!


I haven’t seen this remark by Dennett (who, by the way, is a close friend and co-author of mine — Dan and I joyfully wrote and edited The Mind’s I over 30 years ago!).  I don’t know what I think of it.  I certainly understand the spirit behind it, but I’m not sure whether genuine thinking really depends on the nature of processes all the way down at the level of neurons or even below that.  Of course it’s conceivable that thinking really relies on all that kind of biological hardware at a very fine-grained level, but to me it’s also perfectly conceivable that thinking relies on subcognitive processes at a somewhat higher level, and that the lower levels, such as neurons and neurotransmitters, are just one possible substrate for thinking.  Dennett may have come to the conclusion that neural hardware is absolutely indispensable if one is trying to create a thinking being — but I somehow doubt it.  If he really did say this, though, then I wonder what he would say about the idea of an extremely accurate computer model of a network of a hundred billion neurons, where there was no actual physical neuron, but just billions of complex data structures that triggered each other in subtle ways, acting exactly like neurons but made out of completely different stuff.  I think that if such a model were able to pass the Turing Test, then Dan Dennett (as opposed to John Searle, famous for his “Chinese room” thought experiment, which I think is riddled with flaws, from top to bottom) would be perfectly happy to say that it was carrying out real thinking and that it was conscious and that what it was saying was as filled with meaning as what human beings say.  And so would I!



The End

发布于2014-03-05, 本文版权属于果壳网(guokr.com),禁止转载。如有需要,请联系果壳