Multi-pattern
real-valued spectral associative memories
- Spencer, R.G.
Dept. of Electr. Eng., Texas A&M; Univ., College Station, TX, USA
This paper appears in: Neural Networks, 2001.
Proceedings. IJCNN '01. International Joint Conference on
On page(s): 1180 - 1185 vol.2
15-19 July 2001
Washington, DC, USA
2001
Volume: 2
ISBN: 0-7803-7044-9
IEEE Catalog Number: 01CH37222
Number of Pages: 4 vol. xlvi+3014
References Cited: 7
INSPEC Accession Number: 7024627
Abstract:
A multi-pattern encoding and decoding scheme is presented that
extends the family of spectral associative memories (SAMs) to include
gray-level, or analog patterns. SAMs are frequency-domain formulations
of associative memory that combine the extrinsic redundancy of neural
networks with the in-phase, quadrature, and complex modulation schemes
of communications. Considerable coding gain occurs at the level of
modulation and these networks may be regarded as multi-channel,
multi-carrier generalizations of amplitude modulation. Unlike
multi-pattern bipolar SAMs, which are exclusively content-addressable,
real-valued SAMs also have an addressable mode in which the recall of a
particular memory may be forced. Band structures and anti-aliasing
constraints are presented along with a probabilistic formulation in
which virtual entanglement is a natural feature. Simulations are
presented that demonstrate dual-memory recall for 6x6
gray-level patterns.
Index Terms:
neural nets; content-addressable storage; decoding; encoding;
frequency-domain analysis; singular value decomposition; eigenvalues
and eigenfunctions; probability; spectral associative memories;
multiple-pattern encoding; decoding; frequency-domain analysis;
gray-level; neural networks; singular value decomposition; amplitude
modulation; anti-aliasing constraints; probability; eigenvalues