 |
| Welcome | Projects | Publications | Technical Reports | Software | Resources | Sponsors | Personnel |
| |
| Welcome | Projects | Publications | Technical Reports | Software | Resources | Sponsors | Personnel |
Despite
their superb characteristics, however, neural network based AAMs have
found
limited success in general pattern matching, except as a full frame
adaptive
filter, over the last 50 years. It seems
still there are feature of natural memory which has not been grasped.
One such
feature seems to how we handle dynamic attention.
Ability
of
dynamic search localization is central to many applications where the
principal
challenge is pattern matching. For example, in visual perception, it is
natural
to always focus on some specific objects in the pattern. We never do an
indiscriminate pixel to pixel matching on the entire scene. Similar
localization is needed when there are multiple objects in a scene. In
adaptive
control applications which need to remain functional against sensor
failures at
unpredictable locations, the ability to localize the control response
generation process to the valid segments of input data on the fly is
critical
in order to obtain the best possible control response with respect to
the
information available from the surviving sensors.
It
seems
most main stream AAM and AANs can learn enormous number of patterns
into its
weight set. But, once learned, during retrieval, it requires almost all
the
elements in the pattern space to be present. After the learning is
over,
neither it is possible (no mathematical basis has been shown yet) to
confine
the search dynamically during retrieval within a subset of the
elements. Nor it
is possible to use a tiny fragment of the pattern and obtain the match
with
respect to the specified fragment.
In
this
research we have shown that a novel bi-modal representation of pattern
and a
hologram like complex spherical weight state-space [PsFa95, Gabo69]
offers some
interesting mathematical basis to reproduce the attention aspect of
human
recollection. A number of learning algorithms, including an adaptive
reinforcement learning algorithm can be used in conjunction of this
representation.
In this research we demonstrate a novel holographic system which can adaptively train and learn a massive number of patterns like any other neural network based AAM. However, the difference is that, during retrieval any test pattern or any sub-pattern can be specified as object of focus for retrieval. It is well known, that even the most robust of the conventional AAM ceases to retrieve correctly when the pattern similarity between a stored pattern and the test pattern drops below 70% As we will demonstrate, the proposed memory, with the ability of search localization can overcome this severe limitation. MHAC can retrieve correctly even when the window of focus drops as low as 20%.
This
new
mechanism of dynamic search localization within the general paradigm of
parallel and distributed processing. It now makes the power of
associative
computing available to a new class of pattern matching applications. We
are
investigating its application in various areas such as:
·
Detection
of tiny targets,
·
Background
varying target recognition,
·
Visual
example based content-based image retrieval,
·
Robust
adaptive control systems (which needs to continue operating with small
number
of surviving sensors) in the face of post learning loss of sensors,.
·
Detection
of small irregular patterns (medical diagnostics), will benefit from
this
advancement.
It
promises
direct-visual example based search capability into thousands
(10,000-64,000)
images in near real time (1-10 seconds).
1. Khan Javed. I. & D. Yun, Characteristics
of Holographic Associative Memory in Retrieval with Localizable
Attention, IEEE
Transactions on Neural Networks, Vol-9, Issue-7, May '98, p.389-406.
(GS(6)).
2.
Sutherland,
J. G., A Holographic Model of Memory, Learning
and Expression, Journal of Neural Systems, v1, pp259-267, 1990.
3. A. A. S. Awwal, H. Tang, K. S. Gudmundsson,
Javed I. Khan, Uni-complex and bi-complex representation for
associative memory
with superior retrieval, Proc. SPIE Vol. 4788, Photonic Devices and
Algorithms
for Computing IV, November 2002, pp.159-170.
4. Carpenter, G. A., "Neural Network Models
for Pattern Recognition and Associative Memory", Neural Networks, v.2,
1989.
5. Gabor, D., "Associative Holographic
Memories", IBM J. of Research and Development, 1969, I3, p156-159
6.
7. Javed I. Khan, "Dynamic Sub-Pattern
Matching with Holographic Associative Memory, SPIE proceeding of 27th
Applied
Imagery Pattern Recognition (AIPR) Workshop Dedicated to facilitating
the
interchange of ideas between government, industry, and academia,
Advances in
Computer Assisted Recognition, v. 3584, October 14-16, 1998,
Washington, DC,
pp.174-185.
8. Khan Javed. I. & D. Yun, Holographic
Image Archive, Intl. Journal of Computerized Medical Imaging and
Graphics,
Special Issue on Medical Image Databases, Pergamon Elsvier Science,
U.K., Vol
20 no 4, July-Aug 1996, pp243-257. (OS((1)).
9. Khan Javed.
I., Intermediate Annotationless Content-Based Query and
Retrieval in
Massive Image Archives with Holographic Information Representation ,
Journal of
Visual Communication and Image Representation, Special Issue on
Indexing,
Storage, Retrieval and Browsing of Image and Video, Vol 7, no 4, 1997,
pp378-393. (OS(2)).
10. Khan Javed I & D. Yun, A Parallel
Distributed and Associative Approach for Searching into Image Pattern
with
Complex Dynamics, Journal of Visual Languages and Computing, Special
Issue on
Visual Information Systems, Vol 8, no 2, June 1997, pp303-331. OS((1)).