Biomedical Engineering Reference
In-Depth Information
FUTURE TRENDS
ing concepts (groups of similar patterns), since
it implies the learning of relationships between
patterns mentioned above.
In addition, simulations confirm the idea ex-
pressed in this work, since, by overflowing the
capacity of the net, we can get optimal results of
classification in many cases. This technique is
therefore very useful when tackling classifica-
tion and learning problems, with the advantage
of being unsupervised.
Recently, concept learning has received much
attention from the Artificial Intelligence com-
munity. There are sectors of this community
interested in using techniques such as fuzzy and
classical logics to study this process.
Regarding associative memories and neural
networks, the current trend is to adapt or develop
learning rules in order to increase the capacity
of the network.
FUTURE RESEARCH DIRECTIONS
CONCLUSION
This chapter presents a new open research line,
from the fact that some new modifications of the
learning rule, reinforcing other aspects of the re-
lationships among patterns, may be developed.
An interesting point to be studied is the pos-
sible generalization of Hebb's learning rule such
that the capacity of a neural network increases,
since it is crucial for many applications.
In this work, we have explained that the limita-
tion in capacity of storing patterns in a recurrent
network has not to be considered as determinant,
but it can be used for the unsupervised classifi-
cation of discrete patterns, acting as a concept
learning system.
The neural model MREM, based on multi-
valued neurons, has been developed, as a gener-
alization of the discrete Hopfield model and as
an auto-associative memory, improving some of
the undesirable aspects of the original Hopfield
model: some methods and results for the net not
to store spurious patterns in the learning phase
have been shown, reducing so the number of local
minima of the energy function not associated to
an input pattern.
By applying a slight modification to Hebb's
learning rule, a mechanism to reinforce the learn-
ing of the relationships between different patterns
has been introduced to the first part of the process,
incorporating knowledge corresponding to several
patterns simultaneously. This mechanism can be
repeated iteratively to enhance the relationship
learning procedure.
One of the main facts expressed in this work is
that network capacity, which can be viewed as a
restriction for the network as associative memory,
becomes a powerful ally when forming and learn-
REFERENCES
Athithan, G., & Dasgupta, C. (1997). On the Prob-
lem of Spurious Patterns in Neural Associative
Models , IEEE Transactions on Neural Networks,
vol. 8, no. 6, 1483-1491.
Crick, F., & Mitchinson, G. (1983) . The Function
of Dream Sleep , Nature, vol. 304, 111-114.
Egmont-Petersen, M., de Ridder, D., & Handels,
H. (2002). Image processing with neural net-
works—a review , Pattern Recognition, vol. 35,
2279-2301.
Erdem, M. H., & Ozturk, Y. (1996). A New fam-
ily of Multivalued Networks , Neural Networks
9,6, 979-89.
Hebb, D. O. (1949) . The Organization of Behavior .
New York, Wiley.
Search WWH ::




Custom Search