Re: Finding useful functions- part 1

From: Wolf Kirchmeir (wwolfkir_at_sympatico.ca)
Date: 10/31/04 

Date: Sun, 31 Oct 2004 12:15:23 -0500

Stargazer wrote:
[...]
>
> It seems complicated because we're talking of different paradigms
> here. Under behaviorist parlance, learning is related to the changes
> in behavior of the organism/mechanism that one is studying, while in
> neuroscience and artificial neural networks, learning is a change
> of internal parameters (and/or architecture, if one wants to account
> for plasticity) of the organism/mechanism. Although logically sound
> and coherent, the behaviorist definition prevents the analysis of
> internal modifications of the mechanism that aren't immediately
> reflected in corresponding changes of behavior. For the behaviorist,
> unsupervised learning does not exist, and this may explain why it seems
> difficult to understand it.
>
> *SG*

I don't think that for the behaviorist unsupervised learning does not
exist - as I understand the term, it seems to be more or less synonymous
with classical conditioning, which does not depend on feedback to the
organism's behaviour, but only on repeated presentation of the
conditioning stimulus.

Nor does behaviourism prevent study of the underlying changes; it claims
that that is a different level of analysis, is all. Nor does it claim
that changes in the underlying mechanism must be "immediately reflected
in corresponding changes in behaviour." Drop the "immediately", though,
and you've characterised the behaviourist p.o.v correctly, IMO.

On the whole, I think the distinctions are still fuzzy. According to
Stephen Harris, unsupervised learning may be triggered by external
stimuli, but does not depend on external feedback. Your comments above
and elsewhere imply that unsupervised learning may take place without
external inputs. I would distinguish these as two types of change in the
NN, and call the latter "development."

I earlier said I was uneasy with the distinction between development and
learning, since much development requires external inputs. On
reflection, IMO we should distinguish between:
a) NN changes controlled by external, fed back inputs (supervised
learning, or operant conditioning);
b) NN changes caused by external inputs (unsupervised learning, or
classical conditioning);
c) NN changes cause by changes in the NN itself (??? development???)

In nature, changes in hormones (for example) surrounding the NN may
affect the NN's functions, including those that cause changes in the NN.
Moreover, a NN may be extended, curtailed, or combined with another NN
when the chemistry of the surrounding medium changes. If we avoid
calling such external chemical changes inputs, then we have development
clearly distinguished from learning. Or so it seems to me.
Behaviourally, development would then differ from learning in that it
modifies the repertoire of behaviours that may be shaped, but does not
shape those behaviours. In fact, behaviorism cannot and does not claim
to account for the emergence or production of new behaviours, only for
their shaping by environmental contingencies. For clarity, it should be
said that new combinations of existing behaviors are not not new
behaviours in this sense. Note also that there must be corresponding
development/change in the physiology of the organism - it's no good
producing a NN that controls flight if the organism doesn't have wings
to fly with.

There is of course a complication, in that learning in the behaviorist
sense can be and is influenced by changes in the chemistry surrounding
the NN. So much so, that for, example, if one studies for test while
high on caffeine, one's performance on the test is affected by the
amount of coffee drunk prior to the test. But this is IMO a minor quirk
of the system -- caffeine AFAIK does not trigger NN development as
hormones do.

Relevant Pages

  • Re: Ben G on reinforcement-learning and the wirehead problem
    ... "Reinforcement learning is learning what to do--how to map ... It is of quite unnecessary for the agent to have any opinions about, or evaluations of, any of its behaviours. ... It is merely necessary for the system's architecture to enable conditioning of behaviours. ...
    (comp.ai.philosophy)
  • Re: Gradual Learning, not Reinforcement Learning
    ... You seem to believe that "learning" means acquiring behaviours the system has never exhibited before. ... You use learning HORSE as an example, and posit HOXEL as an intermediate stage, with HORSE as an "ideal" to be achieved. ... Firstly, the system must be capable of "producing language", else it cannot learn the semantic difference between HORSE and HOTEL, both of which are easy consequences of HOXEL. ...
    (comp.ai.philosophy)
  • Re: Cool visual illusion
    ... McLuhan retold a story about showing movies about agricultural methods to some people in sub-Saharan Africa, people who had never seen a movie before They saw things in the movies that the more sophisticated Westerners did not see - eg, a chicken in one corner of a single frame IIRC. ... rewiring that constitutes learning at the neurological level can and does take place at several levels. ... it has to be explained using the biology of the system. ... Most importantly, the system must be capable of some behaviours, else there won't be any behaviours to shape, hence no learning. ...
    (comp.ai.philosophy)
  • Re: What is conditioning?
    ... This is learning, very rapid, but real learning all the same. ... Learning means the shaping and combining of existing behaviours. ... together as it interacts with its environment. ... Skinner exaggerated the emphasis on conditioning kids to become just about anything you wanted to counteract the idea that people were genetically destined to be one thing or another. ...
    (comp.ai.philosophy)