gzz-commits
[Top][All Lists]

## [Gzz-commits] manuscripts/Paper paper.tex

 From: Janne V. Kujala Subject: [Gzz-commits] manuscripts/Paper paper.tex Date: Fri, 28 Mar 2003 07:53:07 -0500

CVSROOT:        /cvsroot/gzz
Module name:    manuscripts
Changes by:     Janne V. Kujala <address@hidden>        03/03/28 07:53:07

Modified files:
Paper          : paper.tex

Log message:
report exper results

CVSWeb URLs:
http://savannah.gnu.org/cgi-bin/viewcvs/gzz/manuscripts/Paper/paper.tex.diff?tr1=1.54&tr2=1.55&r1=text&r2=text

Patches:
Index: manuscripts/Paper/paper.tex
diff -u manuscripts/Paper/paper.tex:1.54 manuscripts/Paper/paper.tex:1.55
--- manuscripts/Paper/paper.tex:1.54    Mon Mar 24 12:24:17 2003
+++ manuscripts/Paper/paper.tex Fri Mar 28 07:53:07 2003
@@ -1242,42 +1242,59 @@
Experiments on black-and-white %(faces,)
ink blots and snow crystals
\cite{goldstein71visualrecognition} show that
-complex pictures can be remembered and recognized and that recognition
+complex, unfamiliar pictures can be remembered and recognized
+and that recognition
performance decreases very little over time.

-We have conducted a pilot experiment (with one subject)
-in a similar setting comparing
-the recognition performance of our textures with solid color backgrounds
-(using colors from the same distribution that is used for the textures).
-First, 15 textures were shown sequentially, 5 seconds each.
-Then, recognition was tested by showing the 15 seen textures
-with 15 unseen textures in a random order, and the subject answered
-old or new. The test was repeated with solid color backgrounds
-for comparison.
-
-The results were as follows:
-for textures, .3 bits per trial of the old/new information got through
-(80 percent correct: 12/15 correctly recognized as seen and
-12/15 correctly identified as not seen);
-for solid colors, only .003 bits per trial got through
-(53 percent correct: 13/15 correctly recognized as seen and
-3/15 correctly identified as not seen). XXX
+We have conducted an experiment in a similar setting comparing
+the recognition performance of our textures with solid color backgrounds.
+Five participants performed the experiment in both conditions
+(in random order).
+
+15 target backgrounds and 15 distractor backgrounds were randomly
+chosen for both conditions.
+The distribution of the solid colors was the same that is used for
+the colors of textures with the difference that the highest
+lighness tail was de-emphasized to increase the otherwise low
+discriminability of very light, unsaturated colors.
+
+First, the 15 target backgrounds were shown sequentially,
+in a random order, 5 seconds each.
+Then, recognition was tested by showing the 15 target backgrounds
+and 15 distractor backgrounds in a random order and
+having the participant answer old'' or new'' for each one.
+The time for answers was not limited.
+
+The results are summarized below
+(percentage of trials and reaction time in seconds):
+\begin{center}
+\begin{tabular}{c|cc|cc}
+      & \multicolumn{2}{c|}{textures \% (RT)}  & \multicolumn{2}{c}{colors \%
(RT)} \\
+      & correct  & incorrect & correct & incorrect \\
+\hline
+old   & 84 (3.0) & 16 (5.6) & 43 (3.4) & 57 (4.0) \\
+new   & 72 (2.6) & 28 (5.5) & 71 (2.8) & 29 (4.1) \\
+\hline
+total & 78 (2.8) & 22 (5.5) & 57 (3.1) & 43 (4.1)
+\end{tabular}
+\end{center}

%80, 53 fail to bring out the results correctly. Should
%use mutual information or something like that. In colors, the subject
%was basically guessing!!!

-Thus, the subject was able to recognize textures ($p < .001$)
-while basically guessing the solid colors.
-Of course, the quantitative results have little statistical significance,
-but in conjunction with our experiences, they do suggest that the
-textures are quite recongizable while solid colors do not have
-enough variation for unambiguous recognition (without additional cues).
+% RT less for correct responses: [F(1,4) = 5.2, p = .082]
+Thus, the recognition performance of the textures was good while
+the solid colors do not have enough variation for unambiguous
+The observed difference in recognition performance is statistically
+significant [F(1,4) = 19.0, p = .012].

-The experiment measured the recognition of only 15 textures,
+Of course, the experiment measured the recognition of only 15 textures,
and surely it would be harder to remember, e.g., 100 textures.
However, the user does not have to remember all the textures;
-it suffices to learn the textures of the most often used documents.
+it suffices to learn the textures of the most often used documents,
+see Fig.~\ref{fig-zipf}.
Many studies of web cache statistics
have shown that file popularity approximately follows Zipf's law
(see, e.g. [xxxref]) so that a small number of documents accounts