|
DEALING WITH DATA VARIABILITY, REPRESENTATIONAL AMBIGUITY AND
SELECTIVE ATTENTION: a Pattern Processing model of perception based on Analog Induction |
Jean-Sylvain Liénard
LIMSI-CNRS
BP 133, 91403 Orsay Cedex, France
e-mail: lienard@limsi.fr
Among the many problems that impede our understanding of perception, three
appear to be of primary importance: data variability, selective attention, and
learning. In the present article we shall examine those three problems and
present a perceptual model, which aims at integrating them into a single
view.
Recognizing a pattern amounts to put many patterns into a single category, and
variability is a result of this classification process. The problem can be
reduced
by associating each object with several categories or attributes, which form a
complete description of it. Thus perception can be viewed as a hierarchical
change
of representation of a constant content. Its goal is to recognize not only the
identity of the object, but also all of its properties that have some perceptual
relevance for the human observer of the same signal or scene.
A computational perception model can be derived from the above considerations.
At the low level we have a representation very close to the signal itself, such
as pixels, spectral values, time-frequency patterns. At the high level we have
abstract properties of the signal, which usually consist of several attributes
specifying the identity, properties and localization of the object in the input
space. Bottom-up functioning of the model consists in activating the appropriate
pixels and obtaining the complete description of the object. Top-down
functioning starts from
a high-level specification and provides a low-level representation in
terms of pixels. In usual situations the information given at both levels is
incomplete, and the aim of the model is to complete both descriptions.
Variability
occurs when several input patterns are associated with the same high-level
description. Similarly, ambiguity occurs when a given low-level pattern can be
associated with several high-level descriptions.
This model provides a framework adequate to study the phenomena of selective
attention. In the presence of a low-level scene comprising many objects, covert
attention is simulated by fixing the values of some high-level descriptors,
which partly define
the properties of the particular object one is interested in. Running
the system yields simultaneous activation at the low level (the pixels
corresponding to the object of interest keep activated, the other ones are
deactivated) and at the high level as well (the values of the descriptors which
were unspecified in the first phase get fixed). Simulation of overt attention
(eye
saccades) may also be simulated with the adjunction of low-level information
representing the eye orientation with respect to the whole visual space.
In our present implementations, dealing with binary data, we develop a
"learning by examples" strategy, that is made possible by the presence of
complete
high-level descriptions. The Analog Induction strategy repetitively uses the
basic
Analogy rule ("A is to B what C is to D"). Thus the system is able to recognize
an
example it has never encountered in this particular position, from the knowledge
of example pairs encountered in both positions.
The principles we implement here may be extended to the study of some aspects
of creativity, in the sense that some new concepts or representations may in
fact come
from a known field and be simply transposed into the new one according to some
transform discovered through a set of analog situations.
e-mail: hmp@ipvvis.unipv.it