
Doctoral thesis of Kyoto City University of Arts
2004
67page(20.54 MB, Japanese)

document video 1
2003.12.05
at Kyoto City University of Arts
2min 59sec(12MB)

document video 2
2004.02.15
at Tokyo International Forum
1min 52sec(10MB)

document video 3
2004.02.15
at Tokyo International Forum
2min 54sec(9.85MB)
|
1. About Shizuka
Shizuka is a kind of artificial intelligence (A.I.) software which the
program draw various images autonomously. This, however, purpose is
not to make a great A.I. artist which can draws better than human beings,
but to understand how to make visual art by human being deeper.
I know this goal would be long way. At this point(-version 3.0), I tried
to realize interactive drawing software, that the program add new images
from audience’s drawing, which means “association memory”.
2. Artificial Intelligence Painter
A.I. painter is not my original concept. It has already been realized
by Harold Cohen, professor at University of California, San Diego,
30 years ago. His program is named AARON. I was, however, keeping
in touch him about this concept then I reached this concept should
become a kind of genre by many researchers and artists, because every
visual representation way is variety in each person. AARON is made
from Cohen’s way to draw. Therefore, AARON can not make images which
Cohen ordered. In this point, I felt that it is more natural that
a program get images itself from outside world. So, my program is
a growing model to represent images which I can also imagine.
3. Cognitive Primitive
Shizuka convert all shapes which audiences drew into special data
format named “cognitive primitive”. (Fig 1) Cognitive primitive
means a shape which doesn’t represent meaning itself though evocate
something to mean to us. Cognitive primitive has not only image
data also has link point to other cognitive primitives. Primitives
have relationship each other. Therefore, the image constructed
with related primitives represent a meaning.

Fig 1: Converting into cognitive primitives
4. Image represents a meaning
We are sharing same images which have a common meaning roughly. A
face, for instance, is drawn two eyes, a nose and a mouse in a
circle. Thinking about this fact, I converted face images into
cognitive primitives (Fig 2). Then, I found there’s a kind of rules
in its relationship.

Fig 2: Relationship of faces
Yet, I found its rule kept its meaning even if shape parts had changed.
(Fig 3)

Fig 3: Examples changed parts in the same rule.
5. Association
Shizuka, theologically, can get infinitive cognitive primitives from
audiences’ drawings. It means there is no limitation to choose.
However, to associate a proper shape, randomized choice doesn’t
succeed always. Then I used Self-Organized Map (SOM) to determine
a similar shape. SOM is one of neural net algorithms proposed by
Talvo Kohonen, professor at Helsinki University of Technology.
It is mainly used in voice recognition, finger-print matching and
so on. SOM can match anything boldly even if the shape is unclear.
Yet, SOM classifies similar shapes to a similar group. With SOM,
if audience draws something, the program pick most similar one
up. And the program can add drawing which evocate surprising meaning
finally. (Fig 4, 5)

Fig 4: Mixing (red part determined by SOM)

Fig 5: Examples of mixing
Figure 6 is a picture made from interaction between an audience
and Shizuka. At first, audience drew just a circle. After then, Shizuka
add drawing autonomously. At last, audience add eyes and mouse to
complete it.

Fig 6: Result
6. Next
Research of Shizuka is not finished yet. There are many topics for
next step. For instance, they’re coloring, compositing and so on.
At least, I will make Shizuka can deal with color next time.
[1] Harold Cohen: http://crca.ucsd.edu/ hcohen/, AARON’s father
[2] Timo Honkela: “Self-Organizing Maps in Natural Language Processing”,
Helsinki University of Technology, Espoo, 1997
|