發信人: GzLi (笑梨), 信區: DataMining
標  題: [合集]ML書里說ANN的研究分兩個團體
發信站: 南京大學小百合站 (Sat Jan 11 21:56:30 2003)

acat (考完了還要干活:() 于Tue Jan  7 22:35:36 2003)
提到：

一個使用之來模擬生物學習過程

一個則脫離生物的過程


第一種的研究是不是太薄弱了。



daniel (飛翔鳥) 于Wed Jan  8 01:02:29 2003)
提到：

  NN is a very complex community. Mitchell just meant two main motivations

of NN researchers. If you peer at the community you may find researchers

with different background are working with different "NN", e.g. these

from mathematics are interested in stability or convergence, these from 

computer science focus on algorithms and generalisation, these

from automation work on neural control, these from electronics investigate

neural circuit, these from physiology want to get illumination from 

artificial NN on biological NN, these from psychology wish explain their

experimental behavior data with NN models, ...
 

【 在 acat (考完了還要干活:() 的大作中提到: 】
: 一個使用之來模擬生物學習過程
: 一個則脫離生物的過程
: 
: 第一種的研究是不是太薄弱了。


strawman (獨上江樓思渺然) 于Wed Jan  8 09:49:21 2003)
提到：

【 在 daniel (飛翔鳥) 的大作中提到: 】
:   NN is a very complex community. Mitchell just meant two main motivations
: of NN researchers. If you peer at the community you may find researchers
: with different background are working with different "NN", e.g. these
: from mathematics are interested in stability or convergence, these from 
: computer science focus on algorithms and generalisation, these
                            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
那咱們的著眼點是不是太淺了些？要設計一個好的算法使其有好的泛化能力，畢竟是要
以數學為基礎的啊，要關心NN的stability, convergence,還有capacity.

再請教：我記得有一篇闡述BP NN 是 universal approximator的文章，不知道在網上
能否找到？能否提示一二？
  
: from automation work on neural control, these from electronics investigate
: neural circuit, these from physiology want to get illumination from 
: artificial NN on biological NN, these from psychology wish explain their
: experimental behavior data with NN models, ...
: 【 在 acat (考完了還要干活:() 的大作中提到: 】


daniel (飛翔鳥) 于Wed Jan  8 12:34:48 2003)
提到：

【 在 strawman (獨上江樓思渺然) 的大作中提到: 】
: 【 在 daniel (飛翔鳥) 的大作中提到: 】
:                             ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
: 那咱們的著眼點是不是太淺了些？要設計一個好的算法使其有好的泛化能力，畢竟是要
: 以數學為基礎的啊，要關心NN的stability, convergence,還有capacity.

I don;t know why you think it is "shallow". They are just different aspects
of an object. Do you think designing algorithms does not require mathematics?
Be aware that nothing can live without mathematics. 

I don't know how do you care the stability and convergence, do you have tried 
to prove the stability or convergency of any practical NN model? For most 
computer scientists, obtaining a well-working algorithm is good enough.
In fact, most properties of most practical NN models have not been proved
at present.

: 再請教：我記得有一篇闡述BP NN 是 universal approximator的文章，不知道在網上
: 能否找到？能否提示一二？

not one. many papers on this topic. one way is from Kolmogrov's work, which 
initials debating for years. The other way is more constructive. But both
are quite difficult to understand for pure computer students. In fact, for
most NN researchers, be aware of the conclusion is enough.
I don't think any of those papers can be found on the internet because they
were worked out at the begining of 1990s.



bigmeat (笑笑生) 于Wed Jan  8 12:48:06 2003)
提到：

【 在 daniel 的大作中提到: 】

: In fact, most properties of most practical NN models have not been proved

: at present.

能不能說具體一點？



strawman (獨上江樓思渺然) 于Wed Jan  8 22:12:03 2003)
提到：

【 在 daniel (飛翔鳥) 的大作中提到: 】
: 【 在 strawman (獨上江樓思渺然) 的大作中提到: 】
: I don;t know why you think it is "shallow". They are just different aspects
: of an object. Do you think designing algorithms does not require mathematics?
: Be aware that nothing can live without mathematics. 
: I don't know how do you care the stability and convergence, do you have tried 
: to prove the stability or convergency of any practical NN model? For most 
: computer scientists, obtaining a well-working algorithm is good enough.
: In fact, most properties of most practical NN models have not been proved
: at present.
     呵呵，我沒有去證明這類的問題。我只是想，利用數學家的這些成果來設計算法。
那既然NN的許多性質并沒有被證明，那么這些算法well-working的本質又是什么呢？
這也許又是數學家的事了。
: not one. many papers on this topic. one way is from Kolmogrov's work, which 
: initials debating for years. The other way is more constructive. But both
: are quite difficult to understand for pure computer students. In fact, for
: most NN researchers, be aware of the conclusion is enough.
: I don't think any of those papers can be found on the internet because they
: were worked out at the begining of 1990s.
: (以下引言省略 ... ...)


yinxucheng (yxc) 于Thu Jan  9 13:35:33 2003)

提到：

我覺得進行應用科學研究與學習，主要有三種方式：


（1）偏理論

主要是利用數學的東西，如定義、定理和證明等；在人工神經網絡的里面，主要指網絡的

數學機理、收斂等。


（2）偏抽象

主要是根據已有數學定理（或沒有），對存在的現象提出相應的模型，然后進行分析，

并設計實驗驗證、改進；在人工神經網絡的里面，主要指網絡模型與結構、學習方法等。

好像，2001年Science上的文章“Machine Learning of Science: State of the Art and

 Future Prospects”就是以這種“偏抽象”的思路來論述的。


（3）偏應用

主要是根據已有的模型，給出需求與分析，設計實現，并應用到具體的實踐中；在人工神

經網絡的里面，主要指針對具體的網絡結構與學習方法，進行算法設計與實現，當然，在

應用的過程中，需要對相應的方法進行改良。像應用BP網絡進行模式識別就是此類方式。


我認為，對于計算機學科（包括其它工科）的學生來說，首先應該是進行“偏應用”的研

究與學習；之后，如果對于某一個具體的點（小方向）很感興趣，可以進行一定的“偏抽

象”學習；至于“偏理論”方式，除非你的數學功底非常好，否則最好不要去嘗試。


不知上面的思路有什么問題，請各位大蝦指點指點！


【 在 acat 的大作中提到: 】

: 一個使用之來模擬生物學習過程

: 一個則脫離生物的過程

: 

: 第一種的研究是不是太薄弱了。