d*****u
发帖数: 17243 | 1 在philosophy of mind领域，一直有connectionism和comptationalism两大传统。
从图灵机理论问世以后，computationalism逐渐成了主流，
但是80年代以后connectionism又重新反扑。
这可能跟神经网络的再度兴起有很大关系。
不知道在人工智能领域有没有一些新的讨论。 |
h*i
发帖数: 3446 | 2 These are on different levels (Marr, 1982): connectionism can be considered
a viewpoint focusing on the algorithm level or lower, computatinalism is
obviously concerns with the computational level.
For AI, computational level is clearly more important, and will probably
remain so in the future, since AI has to be implemented on devices other
than the living brains.
Even for neuroscience, recently, computational level considerations are
gaining attention, e.g. neuroeconomics. The reason is that
【在 d*****u 的大作中提到】
: 在philosophy of mind领域，一直有connectionism和comptationalism两大传统。
: 从图灵机理论问世以后，computationalism逐渐成了主流，
: 但是80年代以后connectionism又重新反扑。
: 这可能跟神经网络的再度兴起有很大关系。
: 不知道在人工智能领域有没有一些新的讨论。
|
d*****u
发帖数: 17243 | 3 但是我觉得David Marr本身就是computationl阵营里的
所以他把computational theory也就是cognition的目标等同于input到output的
mapping
有个很有名的哲学讨论就是Chinese room
一个人通过symbol manipulation完全可以在特定任务中实现“懂中文”的效果
但是他实际不懂中文
所以基于symbol的computation跟思维（或者semantics）有一道鸿沟
而且我觉得人有能力interpret一些parallel distributed的东西
这些东西很难用一个symbol来描述，因为它们永远是那种“似是而非”但又确实存在的
considered
X
【在 h*i 的大作中提到】
: These are on different levels (Marr, 1982): connectionism can be considered
: a viewpoint focusing on the algorithm level or lower, computatinalism is
: obviously concerns with the computational level.
: For AI, computational level is clearly more important, and will probably
: remain so in the future, since AI has to be implemented on devices other
: than the living brains.
: Even for neuroscience, recently, computational level considerations are
: gaining attention, e.g. neuroeconomics. The reason is that
|
D****A
发帖数: 360 | 4 我想知道connectionism和computationalism确切的指什么，有那些主流领域和方向，
解决了那些关键问题，另外根本局限性在哪里。看你们的描述总有些隔岸观火的感觉，
似懂非懂
，哈哈。最好大牛给解答一下。
considered
X
【在 h*i 的大作中提到】
: These are on different levels (Marr, 1982): connectionism can be considered
: a viewpoint focusing on the algorithm level or lower, computatinalism is
: obviously concerns with the computational level.
: For AI, computational level is clearly more important, and will probably
: remain so in the future, since AI has to be implemented on devices other
: than the living brains.
: Even for neuroscience, recently, computational level considerations are
: gaining attention, e.g. neuroeconomics. The reason is that
|
d*****u
发帖数: 17243 | 5 我也在呼唤牛人出现啊
给你转载几点
* Computationalists posit symbolic models that do not resemble underlyin
g brain structure at all, whereas connectionists engage in "low level" model
ing, trying to ensure that their models resemble neurological structures.
* Computationalists generally focus on the structure of explicit symbols
(mental models) and syntactical rules for their internal manipulation, wher
eas connectionists focus on learning from environmental stimuli and storing
this information in a for
【在 D****A 的大作中提到】
: 我想知道connectionism和computationalism确切的指什么，有那些主流领域和方向，
: 解决了那些关键问题，另外根本局限性在哪里。看你们的描述总有些隔岸观火的感觉，
: 似懂非懂
: ，哈哈。最好大牛给解答一下。
:
: considered
: X
|
h*i
发帖数: 3446 | 6 Not much substantially new has been achieved since Marr.
Chinese Room proves nothing. Able to translate a language doesn't mean
knowing the language. From a computational pointview, it's totally different
kind of computation. The point of computational view is to insist on the
primacy of problem formulation: we have to formulate the problems precisely
before talking about them, or it's just hand-waving, wishy-washy.
Nowadays nobody is still taking the symbolic approach to formulate
computational
【在 d*****u 的大作中提到】
: 但是我觉得David Marr本身就是computationl阵营里的
: 所以他把computational theory也就是cognition的目标等同于input到output的
: mapping
: 有个很有名的哲学讨论就是Chinese room
: 一个人通过symbol manipulation完全可以在特定任务中实现“懂中文”的效果
: 但是他实际不懂中文
: 所以基于symbol的computation跟思维（或者semantics）有一道鸿沟
: 而且我觉得人有能力interpret一些parallel distributed的东西
: 这些东西很难用一个symbol来描述，因为它们永远是那种“似是而非”但又确实存在的
:
|
d*****u
发帖数: 17243 | 7 嗯。我不同意你说的这个debate已经不存在了。
只能说在工程领域大家对这个不感兴趣，因为确实不怎么重要。
但是在cogntive science领域，这仍然是话题。
比如在linguistics领域
理论语言学（Chomsky阵营）仍然坚持的是computationalism
到现在为止几乎没变
但是psycholinguistics领域则大量使用connectionist model
像早期的McCellend,还有后来的Seidenberg，Plaut这些人
而且他们明确提出语言学理论可能是错误的，
并不把connectionism看成是一种algorithm implementation
different
precisely
【在 h*i 的大作中提到】
: Not much substantially new has been achieved since Marr.
: Chinese Room proves nothing. Able to translate a language doesn't mean
: knowing the language. From a computational pointview, it's totally different
: kind of computation. The point of computational view is to insist on the
: primacy of problem formulation: we have to formulate the problems precisely
: before talking about them, or it's just hand-waving, wishy-washy.
: Nowadays nobody is still taking the symbolic approach to formulate
: computational
|
D****A
发帖数: 360 | 8 Can I say that nowadays connectionsim has rejuvenated as an adopted child of
computationalism. After years of debate pragmatism has gained its upper
hand.
We've seen rewarding progress in problem solving in many application domains
without better understanding how human brain works. We admit that we came
from a different starting point but we have achieved some plausible results.
However, in the process of pursuing better solutions we sort of opted out
symbolic approaches and picked (statistic
【在 h*i 的大作中提到】
: Not much substantially new has been achieved since Marr.
: Chinese Room proves nothing. Able to translate a language doesn't mean
: knowing the language. From a computational pointview, it's totally different
: kind of computation. The point of computational view is to insist on the
: primacy of problem formulation: we have to formulate the problems precisely
: before talking about them, or it's just hand-waving, wishy-washy.
: Nowadays nobody is still taking the symbolic approach to formulate
: computational
|
D****A
发帖数: 360 | 9 我倾向于把algorithm看成是整个problem solving的building block，而不是单一的
解决方案。所以computationalism和connectionism并不是不可融合的。
另外，我觉得语言学和语言本身是有很大区别的，语言本身肯定不是一个简单的计算
模型，不过computationalism的语言学无可厚非。CS的语言学研究和语言的认知学研
究应该很不一样。
【在 d*****u 的大作中提到】
: 嗯。我不同意你说的这个debate已经不存在了。
: 只能说在工程领域大家对这个不感兴趣，因为确实不怎么重要。
: 但是在cogntive science领域，这仍然是话题。
: 比如在linguistics领域
: 理论语言学（Chomsky阵营）仍然坚持的是computationalism
: 到现在为止几乎没变
: 但是psycholinguistics领域则大量使用connectionist model
: 像早期的McCellend,还有后来的Seidenberg，Plaut这些人
: 而且他们明确提出语言学理论可能是错误的，
: 并不把connectionism看成是一种algorithm implementation
|
h*i
发帖数: 3446 | 10 Where is the debate on this going on in cognitive science? Care to share
some source?
Linguistic is probably lag behind a bit because Chomsky is still alive, but
the mainstream cognitive science has already moved on.
【在 d*****u 的大作中提到】
: 嗯。我不同意你说的这个debate已经不存在了。
: 只能说在工程领域大家对这个不感兴趣，因为确实不怎么重要。
: 但是在cogntive science领域，这仍然是话题。
: 比如在linguistics领域
: 理论语言学（Chomsky阵营）仍然坚持的是computationalism
: 到现在为止几乎没变
: 但是psycholinguistics领域则大量使用connectionist model
: 像早期的McCellend,还有后来的Seidenberg，Plaut这些人
: 而且他们明确提出语言学理论可能是错误的，
: 并不把connectionism看成是一种algorithm implementation
|
|
|
d*****u
发帖数: 17243 | 11 我当然也没法指出谁谁谁在吵这个
但是前一阵看了Christiansen和Chater写的connectionist pscycholinguistics
2001年出版的
里面谈到了大量的争论
真正做研究的人一般不会过于争论这些heuristic的东西
因为没多大意义，也很难发表
但是你从每个人做的东西就能看出差别了
but
【在 h*i 的大作中提到】
: Where is the debate on this going on in cognitive science? Care to share
: some source?
: Linguistic is probably lag behind a bit because Chomsky is still alive, but
: the mainstream cognitive science has already moved on.
|
d*****u
发帖数: 17243 | 12 嗯，其实某些neural network是能模拟一切可计算的函数的（不考虑效率）
所以从计算可能性来说两者一点也不冲突
冲突的主要是哪个能代表认知的实际过程
那些symbol表示的states是不是最好的表征，等等
我还是举个语言的例子吧，因为比较了解一些
英语里有很多词缀很词根，它们都有比较固定的语义和发音
但是有些unit是模棱两可的，比如submit, remit, transmit里的mit
或者grocer, dresser里的er
根据seidenberg等人的观点，当语义和语音有强的connection时，
就产生了词缀的概念
当有相对较弱的connection时，就产生了这些模棱两可的情况
这些都是可以计算的，也可以通过实验证实
所以他说词缀的概念整个就是表象，实质是connection
【在 D****A 的大作中提到】
: 我倾向于把algorithm看成是整个problem solving的building block，而不是单一的
: 解决方案。所以computationalism和connectionism并不是不可融合的。
: 另外，我觉得语言学和语言本身是有很大区别的，语言本身肯定不是一个简单的计算
: 模型，不过computationalism的语言学无可厚非。CS的语言学研究和语言的认知学研
: 究应该很不一样。
|
D****A
发帖数: 360 | 13 makes sense. 对computationalist来说，他们关注的状态空间是抽象问题的状态空间
跟认知学者关注的不一样，所以其实很难比较好坏。
1. 可计算性本身是不考虑效率的
2. NN是个一般的模型，不过用计算机模拟的NN计算能力不会超过可计算函数
还是，模型是经验的模型，本质上是超越可计算的
之所以你说“可以计算出来”是因为实现中用了模拟和近似
不过我不觉得有什么“智能”是模拟不出来的 |
h*i
发帖数: 3446 | 14 Nick Chater is one of the most active Bayesian probabilistic guy, who's
firmly committed to the computational viewpoint. I would assume the
connectionist approach was taken as no more than an implementation level
alternative to symbolic approach, which is not surprising at all.
Although the ruled based symbolic approach to natural language is commonly
considered outdated in the research circle, in large scale industrial
applications, well designed rule based approach actually works better than
【在 d*****u 的大作中提到】
: 我当然也没法指出谁谁谁在吵这个
: 但是前一阵看了Christiansen和Chater写的connectionist pscycholinguistics
: 2001年出版的
: 里面谈到了大量的争论
: 真正做研究的人一般不会过于争论这些heuristic的东西
: 因为没多大意义，也很难发表
: 但是你从每个人做的东西就能看出差别了
:
: but
|
h*i
发帖数: 3446 | 15 What you said seems to be consist with my view, which can be summarized as "
mental states are probability distributions defined on mental configuration
space."
Symbolic representation of states can be considered as the abstraction of
structures in mental configuration space. The structures are real, but only
in the probabilistic sense.
【在 d*****u 的大作中提到】
: 嗯，其实某些neural network是能模拟一切可计算的函数的（不考虑效率）
: 所以从计算可能性来说两者一点也不冲突
: 冲突的主要是哪个能代表认知的实际过程
: 那些symbol表示的states是不是最好的表征，等等
: 我还是举个语言的例子吧，因为比较了解一些
: 英语里有很多词缀很词根，它们都有比较固定的语义和发音
: 但是有些unit是模棱两可的，比如submit, remit, transmit里的mit
: 或者grocer, dresser里的er
: 根据seidenberg等人的观点，当语义和语音有强的connection时，
: 就产生了词缀的概念
|
h*i
发帖数: 3446 | 16 Actually, there are cognitive scientists who takes the abstract
configuration space of a problem as the model of mental space, e.g. Roger
Shepard. I personally think that is the right view. Even if the actually
mental space is different, the difference has to be understood in the
framework of abstract configuration space.
【在 D****A 的大作中提到】
: makes sense. 对computationalist来说，他们关注的状态空间是抽象问题的状态空间
: 跟认知学者关注的不一样，所以其实很难比较好坏。
: 1. 可计算性本身是不考虑效率的
: 2. NN是个一般的模型，不过用计算机模拟的NN计算能力不会超过可计算函数
: 还是，模型是经验的模型，本质上是超越可计算的
: 之所以你说“可以计算出来”是因为实现中用了模拟和近似
: 不过我不觉得有什么“智能”是模拟不出来的
|
