Page 372 - Introduction to Information Optics

7.1. Basic Architectures

Square Law
^ / Detector
ki
Spatial domain—>T>
filter h(-x.-y) ^ 1st STEP
N
Fourier
Spatial Domain
Domain

X
x X
N X
X X N 2nd STEP
Fourier Output
Domain Plane

Fig. 7.2. A joint-transform correlator (JTC).

7.1.2. NEURAL NETWORKS

Digital computers can solve some classes of computational problems more
efficiently than the human brain. However, for cognitive tasks, such as pattern
recognition, a 3-year-old child can perform better than a computer. This task
of pattern recognition is still beyond the reach of modern digital computers.
There has been much interest in the implementation of associative memory
using optics, much of it centered around the implementation of neural
networks (NNs). The associative memory is a process in which the presence of
a complete or partial input pattern directly results in a predetermined output
pattern. The neural network shown in Fig. 7.3 is one of the possible implemen-
tations for associative memory. A neural network attempts to mimic the real
structure and function of neurons, as shown in Fig. 7.4. However, it is not
necessary to exactly mimic a human brain in order to design a special-purpose
cognitive machine, just as it is not necessary to exactly imitate a bird in order
to make a flying machine. In view of the associative-based neuron model of
Fig. 7.4, it is, in fact, a matrix^vector multiplier for which a single-layer NN
can be implemented by optics, as will be shown later.

7.1.3.

It is apparent that a purely optical system has drawbacks which make
certain tasks difficult or impossible to perform. The first problem is that optical