The picture above shows graphically some of the results of visual analysis done on the given input (the parallelogram), after pressing on the "Start" button. Notice that the picture is captured from an earlier version of the program. It has not been updated since because the visual analysis is now more complex (more features are perceived and shown graphically), and would be harder to explain. Also notice that the analysis was interrupted (by clicking on the Pause button) at an early stage, roughly 1/2 second after it started.

Yellow lines show the approximation of "median lines" done by the program on the thick, hand-drawn actual lines. At this early stage, three of the yellow median lines are continuous, while the fourth one is dotted: this is a rough way to show the "certainty" with which Phaeaco created those lines. Further analyis would add more points to the fourth line, and would make it appear continuous, too.

Intersection points, or rather, *vertex* points in this
case, are shown with a small inverted colored triangle. In other
cases, a small square depicts *touch* points (as the one
formed by the two lines of letter T), and a small diamond depicts
*cross* points (as in letter X). More complex
intersections (as in letters K and Y) are depicted with a small
star. The color of the intersection point shows the certainty
with which the program sees this as such a point. This certainty
is directly related to the certainty of the lines forming the
intersection. For example, the two vertices on the right side of
the parallelogram are shown with red color (lowest certainty)
because the program is still uncertain about the constituent
right side (dotted line). The other two vertices on the left have
yellow and green colors because the program is more certain about
their constituent lines. The colors used for intersection points
are red, orange, yellow, green, cyan, and blue, in increasing
order of certainty.

Other visual compenents that can be depicted graphically on
the input are: *border lines,* when the "black"
portion of the input is not sufficiently linear, and thus defines
a filled (black) area, *cuves, areas* (either filled, or
outlined ones), and more.

It should be noted that all the above color-coded components provide just a superficial glimpse over the progress of visual analysis. A better report can be obtained at the end of processing, part of which is shown on the right page of the notebook, above. The information on the right shows that Phaeaco found the following features and relations:

- "# of components: 1 (occurs 1)": The whole input consists of one connected component (one parallelogram). If there were one parallelogram and one (disjoint) triangle, two components would be observed.
- "# of lines: 4 (occurs 1)": There are four line
segments in the input. The phrase "occurs 1" in
parenthesis shows that the percept "four lines"
occurs once in the input. If two quadrilaterals were
drawn (e.g., a parallelogram and a square), this percept
would be reported as follows: "# of lines: 4 (occurs
2)", i.e., we would have
*twice four-lines*in the input. In the case of one parallelogram and one triangle we would have: "# of lines: 4 (occurs 1)", and on the next line "# of lines: 3 (occurs 1)". - "# of vertices: 3 (occurs 1)": There are three vertices. We know there should be four, and actually the program has identified all four of them as shown by the low-level colored indicators on the picture. However, at the stage of processing at which the program was interrupted (less than one second after it started) the fourth vertex did not manage to be observed explicitly and be associated with a node in the internal representational network. (It is the contents of this network that we see on the right page.) Given a few more tenths of a second, all four vertices would be observed and reported.
- "Equal lengths: 2 (occurs 2)": There are two
equal lengths, and this percept occurs
*twice*. In the case of a square the percept would be: "equal lengths: 4 (occurs 1)", reporting the four equal sides of the square. - "Equal slopes: 2 (occurs 2)": Similarly, there are two equal slopes, and this percept occurs twice: two parallel horizontal sides, and two parallel tilted sides. We see that this percept does not care about the actual slopes; all it registers is the existence of parallelism. If there was a third horizontal line dividing the parallelogram from left side to right side, we would get: "equal slopes: 3 (occurs 1)", for the three horizontal sides, and on the next line "equal slopes: 2 (occurs 1)" for the two tilted sides.
- "Equal angles: 2 (occurs 2)": Finally, the percept of "two equal angles" occurs twice. In the case of a rectangle, or square, the percept would be: "equal angles: 4 (occurs 1)", reporting the four right angles.

It should be noted that Phaeaco is not as good at *counting*
as the above description might suggest. When the numbers are
small (up to about four or five), counting is generally accurate.
However, as the actual number of occurrences (of various
percepts) increases, Phaeaco's reporting of quantities (the
"numerosity" in psychological jargon) becomes
increasingly blurry and inaccurate. For example, if ten straight
lines exist in the input, the lowest level line-detecting
processes will actually identify all of them, but when the time
comes to estimate and report their number Phaeaco will register a
statistical quantity with mean value of around ten, and a certain
standard deviation. The larger the actual number, the larger the
standard deviation. If Phaeaco is asked to report a specific
number, it may report any number around the mean, allowed by the
deviation. What happens in the report shown on the picture above
is that the deviation for numbers up to five is almost zero, so
the exact number is reported.

Last update: 04/14/00

Back to Mentor Section