Page 429 - Introduction to Information Optics
P. 429
414 7. Pattern Recognition with Optics
(b)
Fig. 7.49. (a) Arabic-Chinese numerics training sets, (b) Positive (left sides) and negative (right
sides) parts of the heteroassociation IWM for Chinese to Arabic numerics translation, (c) Positive
(left sides) and negative (right sides) parts of the HA IWM for Arabic to Chinese, (d), (e), Partial
input patterns to the corresponding input patterns to the corresponding output translations.
and Chinese numeric numbers are presented to the NN. Their heteroassoci-
ation memory matrices for translating Arabic to Chinese numeric numbers and
Chinese to Arabic numeric numbers are shown in Fig. 7.49. The corresponding
translated results from Chinese to Arabic and Arabic to Chinese numerics for
partial input patterns are shown in the figure. Thus, we see that NN is
inherently capable of retrieving noisy or distorted patterns. In other words, as
long as the presented input pattern contains the main feature of the supervised
training patterns, the artificial NN has the ability of retrieving the pattern (in
our case the translated pattern), as does a human brain.
7.8.2. RECOGNITION BY UNSUPERVISED LEARNING
In contrast with supervised learning, unsupervised learning (also called
self-learning) means that the students learn by themselves, relying on some
simple learning rules and past experiences. For an artificial NN, only the input