Posted by Reilly Atkinson on January 23, 2000 at 01:59:10:

In Reply to: Knowledge acquisition & learning posted by Denham on January 22, 2000 at 16:00:36:

Denham -- That's all well and good. But the plain fact is that this artificial distinction and others made by the AI mavens helped erode the reputation of AI and helped to sabotage the chances of AI in the commercial market. I saw Kendell(AI) square go boom to bust quite rapidly, and a lot of LISP cats were then on the street.

I say artificial simply because I've built models that qualify in both categories -- these involved heuristics and had adaptive characteristics. As most any practical & experienced model builder will tell you: it's rather dumb to build the "KA" type models by themselves. You use "Learning" or adaptive techniques to make the model better, more powerful, keep it updated, make changes, and so forth. In practice, at least in my experience, model builders make no such distinctions.

And, by the way, a very strong case can be made against the standard claim that neural networks learn. The problem is that neural networks are simply a particular form of non-linear models, and the whole process of finding weights is nothing more than "curve" fitting. Generally, the ability of networks to generalize is like a set of measure zero in the search/event space. Typically, we do not consider a linear regression fit to be learning, so, one may reasonably ask, why is a non-linear fit a learning process?

You say that your distinction is important. I've been in the model building business for almost 40 years. I can see no practical value to your distinction, nor have I encountered it in real life. What's the value?
Regards,
Reilly

---

Follow Ups:

• On 'Artificial' and 'Real' Learning Yogesh Malhotra 19:07:49 01/23/00 (2)
  • Re: On 'Artificial' and 'Real' Learning Reilly Atkinson 16:11:55 01/25/00 (1)
    • Re: On 'Artificial' and 'Real' Learning Yogesh Malhotra 22:05:49 01/25/00 (0)

---

Click Here to Post Follow Up in New Forums

    Knowledge Management Think Tank (New)

  

 Subject: 

Message: