🚀 go-pugleaf

In article <343d5a33.43513184@news.earthlink.net>
           savainl@earthlink.net "Louis Savain" writes:

> In article <876306121snz@longley.demon.co.uk>,
> David@longley.demon.co.uk (David Longley) wrote:
> 
> > The Rescorla-
> >Wagner model of 1971 provided the first good summary of the  data 
> >covering "blocking", and Rescorla's earlier work in 1968  showing 
> >that  the CS needs to bear an "informational", more accurately  a 
> >contingent, temporal relationship to the UCS. Simultaneous events 
> >on   the  other  hand  result  in  overshadowing,  and   backward 
> >conditioning  very  little if anything.
> 
>   I've acknowledged the importance of work in operant conditioning.
> This work however sheds no light on the actual mechanism at the single
> neuron level.  Without a knowledge of this mechanism, no progress can
> be made.

The above is classical conditioning, and there is plenty of  work 
within  this  tradition  going back decades.  Some  of  the  most 
interesting is that of Richard Thompson's group.

> 
> > But a  temporal  relation 
> >alone  is not sufficient (as Kamin showed long ago). If one  goes 
> >beyond  "delay",  to "trace" conditioning, it looks  like  higher 
> >levels of processing such as the hippocampus is required - but as 
> >I've said before, this is already getting into levels of learning 
> >in the neuroaxis...
> 
>   I'm sorry, this is nonsense.  A signal is but message indicating
> that something happened at a specific time.  The neuron processing the
> signal has no idea what that something is and could not care less.  It
> only gets the signal and that is it.  To make sense of anything it
> needs to find several signals to determine their relationships, if
> any.  Temporal relationship is the **only** type of relationship that
> can exist between events (signals).  Period.  David, you are falling
> headlong into intensional superstition here and your appeal to
> obsolete authority makes it worst.

*All* of classical conditioning is based on temporal order.  What 
I'm  talking  about  above is  different  classical  conditioning 
paradigms, ie relations between CS-UCS delivery. 

> 
> >What  I see here in these threads is some excitement on the  part 
> >of  proponents  of  this work, but it seems  to  reflect  "their" 
> >personal  discovery  of something which a large number  of  other 
> >researchers have been working on in the context of small  systems 
> >models  (Richard Thompson's group) for several decades now.  That 
> >work  makes no claims to being able to build "learning  machines" 
> >on the other hand, just unravel some of the neuromechanics behind 
> >basic conditioning processes.
> 
>   The actual physiological mechanism responsible for temporal
> processing is only now making it to the surface (see the above
> reference).  All we had before that was behavioral studies in
> classical conditioning.  Not much more.  I do agree however that
> classical conditioning tells a powerful tale about biological learning
> but we are pragmatic technologists and we need hard *mechanisms* to
> work with not just observations of behavior.

That's  not  true as I have said. Go back and  look  through  the 
neuroscience literature. Look at the work on LTP, and  especially 
look at Thompson's work.

> 
> >To be a breakthrough, a discovery has to be more than one getting 
> >excited   about   something  one  learns  -  to  be   a   genuine 
> >breakthrough, one needs to know the literature and be working  in 
> >the field to know what is and is not new and worthy of attention. 
> >How else can one ascertain whether it is indeed a breakthrough?
> 
>   Sorry, you have no idea what you are talking about in the context of
> recent work done in neuroscience.

Having  spent  4  years in the Division  of  Neurophysiology  and 
Neuropharmacology  working on the neural and neurochemical  basis 
of learning as a postgrad I can assure you that I know  something 
about neuroscience. 

> 
> >If you want to make any real progress on "learning machines"  you 
> >are  going  to  have  to come to grips with  the  issues  I  have 
> >discussed elsewhere, 
> 
>   David, your fascination is with deduction and that has absolutely
> nothing to do with learning.  So you could not possibly understand
> anything worthwhile about the subject.  Learning is forcibly inductive
> and this is the pure extensional truth whether you like it or not.

You  have a preconception that this is the case. What  you  don't 
appreciate  is  why  I  have  spent  so  much  time  and   effort 
critcisiing these preconceptions. I was doing so long before  you 
starting making your "announcements".

But  as I say - let's see something other than - "I've read  this 
article"  and "we have a breakthrough" (but haven't got  anything 
to show for it yet - just a conviction).

Show me (and everyone else) that I'm wrong....

> 
> >the most important of which is that most  of 
> >our psychological idioms, including those of learning and memory, 
> >are  intensional idioms, not the stuff a science  and  technology 
> >can or should be grounded upon (and to fully appreciate that, you 
> >need to know the literature).
> 
>   More nonsense.  Work in computational science is not based on
> intensional idioms but on hard scientific data.  Por favor David, do a
> little research in this area before offering your opinion or
> criticism.  In conclusion I'll just repeat a powerful extensional
> truth that seems to escape you and others:

As  I  say,  I've worked on the problem of the  neural  basis  of 
learning. That's where a lot of my criticism comes from. Too many 
folk  in neuroscience have just uncritically imported ideas  from 
psychology without understanding very much about their origin.

With  all  of the models of learning which have  been  programmed 
since  the  late 60s (and the implications of  the  Church-Turing 
Thesis), why do you think there aren't ANNs running machinery  or 
missiles  all  over the place? - Might it not be  because  what's 
required  are reliable, predicable machines - not ones which  are 
at  the  mercy of local characteristics of the  environment.

But go ahead - "build" your machine, and find out for yourself.
-- 
David Longley (check end reply line #)

Longley Consulting                                                  London, UK
Behaviour Assessment & Profiling Technology,
Research, Data Analysis and Training Services,
Small IT Systems                                http://www.longley.demon.co.uk