No subject

<a-dimitrov at uchicago.edu>

First, I'd like to point out that hardly anybody considers
instanteneous learning in the way you define it. The state of any
network will depend on the environment to which it was exposed. I'd
say that from what you discuss on cannot discrimintae between
real-time and delayed learning unless some more analysis is
done. Since the two groups performed different tasks after
the test, of course they will end up with different
networks, memory or not. So in particular, it does not mean
that.

> all persons trained in that particular motor task should have had 
>more or less the same "trained net," performance-wise, at the end 
>of that training session, regardless of what they did 
>subsequently.

> What are the real implications of this study? One of the most
> important facts is that although both groups had identical 
>training sessions, they had different levels of learning of the 
>motor task because of what they did subsequent to practice. From 
>this fact alone one can conclude with some degree of certainty 
>that real-time, instantaneous learning is not used for learning 
>motor skills. How can one say that? One can make that conclusion 
>because if real-time learning was used, there would have been 
>continuous and instantaneous adjustment of the synaptic strengths 
>or connection weights during practice in whatever net the brain 
>was using to learn the motor task. This means that all persons 
>trained in that particular motor task should have had more or less 
>the same "trained net," performance-wise, at the end of that 
>training session, regardless of what they did subsequently. (It is 
>assumed here that the task was learnable, given enough practice, 
>and that both groups had enough practice.) With complete, 
>permanent learning (weight-adjustments) from "real-time learning," 
>there should have been no substantial differences in the learnt 
>skill between the two groups resulting from any activity 
>subsequent to practice. But this study demonstrates the opposite, 
>that there were differences in the learnt skill simply because of 
>the nature of subsequent activity. So real-time, instantaneous and 
>permanent weight-adjustment (real-time learning) is contradictory 
>to the results here.  


I'd like to disagree with you again. There are numerous examples of
networks without explicit memory (e.g. any BP net), which
generalize pretty well. This is a consequence of their
general approximator property.

> One has to remember that the essence of learning is 
>generalization. In order to generalize well, one has to look at 
>the whole body of information relevant to a problem, not just bits 
>and pieces of the information at a time as in real-time learning. 
>So the argument against real-time learning is simple: one cannot 
>learn (generalize) unless one knows what is there to learn 
>(generalize). One finds out what is there to learn  (generalize) 
>by collecting and storing information about the problem. In other 
>words, no system,  biological or otherwise, can prepare itself to 
>learn (generalize) without having any information about what is to 
>be learnt (generalized).
> 


I don't think anyone has claimed otherwise. We all agree it is most
likely a statistical process and it needs many examples for
learning.

--------------------------------------------------------