Classical Neural Networks - Neural Networks and Micromechanics - page 18

visual cortex, and C-cells resemble complex cells. The layers of S- and C-cells are
arranged alternately in a hierarchical structure. S-cells feature extracting cells and
have variable input connections, which can be modifiable during training. C-cells
have fixed, non-modifiable input connections.
In recent years, these ideas have found new applications [ 65 - 67 ], and [ 68 ].
In Japan, at Toyota Central R&D Labs., Inc., the neocognitron is used for position
detection and vehicle recognition [ 65 ]. This system is tolerant to deformations and
shifts in the position of a vehicle. K. Fukushima [ 66 ] applies the neocognitron for
handwritten digit recognition, but several new ideas have been introduced, such as
the inhibitory surround in the connections from S-cells to C-cells. Fukushima also
applies the neocognitron to the recognition of patterns that are partly occluded [ 68 ].
Other authors [ 67 ] have modified the neocognitron and used it for breast cancer
detection. The so-called Shape Cognitron (S-Cognitron) is composed of two mod-
ules and was introduced to classify clustered micro calcifications, which generally
present an early sign of breast cancer. The first module serves as a shape orientation
layer and converts first-order shape orientations into numeric values. The second
module is made up of a feature formation layer followed by a probabilistic neural-
network-based classification layer. This system was tested on the 40-mammogram
database provided by the Department of Radiology at the University of Hospital
Nijmegen in the Netherlands and showed promising results.
2.6.5 Backpropagation
The backpropagation method was introduced by Paul Werbos in 1974 [ 46 ]. In
1985 -1986, D. Rumelhart and others worked out and applied this mechanism for
neural network training [ 44 , 45 ].
At present, the backpropagation method is used to resolve different practical
tasks. For example, it is used for retrieval algorithms for geophysical parameters in
the retrieval of atmospheric water vapor and cloud liquid water content over oceans
from brightness temperatures measured by the multi-frequency scanning micro-
wave radiometer launched onboard satellite [ 69 ]. These studies have demonstrated
the great potential of neural networks in a large variety of remote sensing and
meteorological applications. For this concrete task, the multilayer perceptron
(MLP) with backpropagation training algorithm was used. MLP has the ability to
detect multiple nonlinear correlations from the training database. MLP has advan-
tages over statistical regression methods [ 70 ].
The backpropagation method is used very often in medicine, for example, for
classifying balance disorders using simulated sensory deficiency [ 71 ]. This task is
important for medical rehabilitation of patients with sensory deficiency. Another
example is connected with the classification of neck movement patterns related to
Whiplash-associated disorders (WADs) using a resilient backpropagation neural
network [ 72 ]. WADs are a common diagnosis after neck trauma, typically caused
by rear-end car accidents. Neck movement was used as input. Rotation angle and