[From Bruce Abbott (960923.1740 EST)]
I have a question that relates to the old concept called "knowledge of
results." You are attempting to learn a difficult discrimination: you are
trying to detect a slight change in resistance to airflow as you breathe air
through a narrow tube. The experimenter imposes a small restriction (or
not; you don't know which) that slightly increases airflow resistance. You
then try to sense whether the restriction has been imposed or not and report
your decision to the experimenter. The experimenter then tells you whether
your decision was correct or incorrect (knowledge of results). What happens
is that, over trials, you get better at detecting the restriction. This
effect does not occur when you are not given knowledge of results.
If we assume that you, the participant, are trying to detect the restriction
as you have been asked to do, then one way to look at your performance is to
assume that you are controlling for saying "present" when the restriction is
present and "absent" when the restriction is absent. You have certain
perceptual signals available to you in the form of internal sensations
(mostly related to muscular effort of the diaphram and muscles of the chest
wall), but it is unclear to you which sensations and at what levels indicate
the presence of a restriction as compared to absence. In the absence of
knowledge of results, you make a guess based on the available input and your
current perceptual function, a guess that may be correct or incorrect.
However, the size of the restriction has been carefully selected, based on
the outcomes of a series of screening trials, so that you are basically
guessing much of the time. It is not clear to you whether your decision is
correct or not.
When you are given knowledge of results, you know when your decisions are
correct or incorrect, but you do not necessarily know on what basis you
arrived at the correct decision when you do. As the trials go on, it
becomes clearer to you when the restriction is present or absent, and you
may be able to describe at least to some degree just what aspect of the
input you are paying attention to when making the decision. My question is,
how has knowledge of results contributed to this improvement in perception?
Regards,
Bruce
[From Shannon Williams (960924.01:00 CST)]
From Bruce Abbott (960923.1740 EST)-
When you are given knowledge of results, you know when your decisions are
correct or incorrect, but you do not necessarily know on what basis you
arrived at the correct decision when you do. As the trials go on, it
becomes clearer to you when the restriction is present or absent, and you
may be able to describe at least to some degree just what aspect of the
input you are paying attention to when making the decision. My question is,
how has knowledge of results contributed to this improvement in perception?
Thinking in terms of neural nets, I think
that the process that you describe should
go something like this:
Let there be two neural nets which are set
up as
input data --> [net1] --> [net2] --> output
Net1 learns to output some pattern based on
whether the restriction is present (or based
on whether the restriction is not present).
Net2 learns to recognize when Net1 is outputing
its pattern, and indicates an appropriate output
for present or not present.
The process would be something like:
1) enter input data
2) based on the output of the net2, make a guess.
3) if you guess wrong, re-weight net1.
if you guess right, re-weight net2 (so that
your output signal is more certain).
The result of this process is that net1 learns to
filter the input data, and net2 learns to read the
filtered data. This results in improved
perception.
···
--
Shannon Williams
sdollars@airmail.net
http://web2.airmail.net/sdollars/shannon.html
[Jeffrey B. Vancouver 092496.1435]
[From Bruce Abbott (960923.1740 EST)]
I have a question that relates to the old concept called "knowledge of
results." etc.
I think you are describing the process of reorganizing the input function
for a goal that is "to make the correct response." There is no input
into the control system that turns the reorganizing process on or off
with knowledge of results. I also think that Shannon said the same thing
using neural net language.
Later
Jeff
BTW, I am very backlog in looking at the net. K of R just happened to
catch my eye.
[From Rick Marken (960924.1300)]
Jeffrey B. Vancouver (092496.1435) to Bruce Abbott (960923.1740 EST) re: KR
I think you are describing the process of reorganizing the input function
for a goal that is "to make the correct response." There is no input
into the control system that turns the reorganizing process on or off
with knowledge of results.
It couldn't have been said better! EXCELLENT Jeff!
Best
Rick