The Best of Creative Computing Volume 1 (published 1976)
Computer-Based Experiments in Cognitive Psychology (Dartmouth College's project COMPUTe, experiments, pattern recognition,)
Computer-Based Experiments in Cognitive Psychology
William L. Bewley
Computer Technology Program
Northwest Regional Educational Laboratory
Portland, Oregon 97204
Author is formerly of Dept. of
Psychology at Lawrence University
in Appleton, Wisconsin
EXPERIMENT 1: PATTERN RECOGNITION
This paper describes the first of six computer-based
experiments contained in Cognitive Psychology: A
Computer-Oriented Laboratory Manual, a product of
Dartmouth College's project COMPUTe. The six
experiments were designed for use in any introductory
or advanced undergraduate psychology course either
partially or totally concerned with human cognition.
Each experiment is run on a time-sharing system
using a terminal as the input/output device.
The computer programs which run the experiments
require 8K storage and are written in Dartmouth
BASIC. Versions which run on DEC'S RSTS
system are also available, and the programs are
presently being modified to run in HP BASIC.
Each program performs two functions: (1) it
runs the student as a subject in an experimental
task, e.g., visual search, continuous memory,
paired-associate learning, concept learning, a game
similar to the Prisoner's Dilemna, and the Missionaries
and Cannibals problem; (2) it runs a simulation of an
information-processing model on the task,
e.g., Pandemonium, the Atkinson and Shiffrin
buffer model, Hintzman's discrimination-net model,
Levine's focusing model, a modification of
Messick's social motives model, and the Newell and
Simon General Problem Solver. The student is
asked to compare his performance on the task with
that of the model and to compare what he thinks
he did in the task with what the model says he did.
The programs are fairly flexible in that some
features of the task and model can be changed by
the student so that he can run his own experiments
on other students and the model.
Background
According to the Pandemonium model of
Selfridge (1959), pattern recognition is a hierarchical
process in which information from input patterns is fed
through a succession of analyzers or "demons," the output
of one demon serving as the input of the next. At the
lowest level are image demons which simply form an image
or unprocessed copy of the input stimulus. This image is
then processed by a set of feature extraction
demons which, as their name implies, extract from
the image such features as straight lines, curved
lines and angles. Cognitive demons then evaluate
the extracted features to determine the likelihood
that one of the several alternative patterns is the
input pattern. These likelihood estimates are then
sent to a decision demon which makes the recognition
response, deciding that the most likely alternative
is the input pattern. There are two strong
implications of this model. The first is that there
are hierarchical levels of processing in pattern
recognition. The second is that the processing at
each level is parallel, i.e., all image demons process
the input pattern simultaneously, all feature
extraction demons process the images simultaneously,
and all cognitive demons evaluate the extracted
features simultaneously. The purpose of this experiment
was to test these two implications.
The Experiment
This experiment is a partial replication of the
visual search experiment of Neisser (1963). Each
student is shown 20 lists of letters, one list at a
time. Each list is composed of 50 lines, 6 letters
per line, arranged in 5 rows with 10 lines per row.
The student's task is to search each list for a
particular target letter, scanning the list from left
to right within each row starting with the top row
and working down (i.e., the scanning pattern used
in reading). A question mark is typed immediately
below each list to indicate that the student may
begin his search. When he is ready to begin he
presses the carriage return button again. The
computer then prints the line number at which the
target occurred and the student's search time in
seconds. The simulation of the Pandemonium
model then scans the same list for the same target,
and its search time, in arbitrary units, is printed.
Following this, the computer types the next list.
This procedure is repeated until a search time is
obtained for each of the 20 lists.
There are two independent variables: the
target to be found and the context in which it is
embedded. The five targets are Q, Z, not-Q, not-Z,
and Q or Z. If the target is Q, only one of the 50
lines in the list contains a Q. If the target is Z, only
one of the 50 lines contains a Z. If the target is
not-Q, only one of the 50 lines does not contain a
Q. If the target is not-Z, only one of the 50 lines
does not contain a Z. lf the target is Q or Z, one of
the 50 lines contains either a Q or a Z. ln every
case, the subject is searching for the line containing
the target. The context in which the target is
embedded is either angular or round. For the
angular.context, the non-target letters are drawn
from the set E, I, M, V, W and X. For the round
context, the non-target letters are drawn from the
set C, D, G, O, R and U. Since there are five targets
and two contexts, there are ten experimental
conditions. The 20 lists shown to each subject
represent only one experimental condition. That is,
the 20 lists have the same target and context. The
only difference among the lists presented to each
subject is the line at which the target occurs.
