Neural Networks With Learning Disabilities
by Connie Brown
Neural networks are systems that use architecture similar to the brain as a means to analyze data and
find hidden relationships. A neural network is well suited for the task of pattern recognition and
predicting trends. To this end, the neural network needs the information in a particular format;
otherwise, learning will not occur. Here, Connie Brown of Elliott Wave International describes learning
problems that neural networks may have and discusses common causes and solutions to help put your
neural network back on track.
During the training phase of a neural network, the system spends extensive time trying to decide what
input information is important to contributing to the output forecast or prediction. It is natural to want to
feed as much input into the neural network as is required to solve the problem, but the information you
need to solve a problem may not be the format or information your neural network needs.
Say that your neural network is about to be set in action for the first time to find the hidden relationships
from your data to predict future prices of your selected market. Anticipation grows. The new spreadsheet
represents endless hours of collecting, cleaning and normalizing data of various indicators. But after
extensive training time, the number of incorrect predictions, or the neural network error rate, does not
decline. The network appears to be unable to learn. Figure 1 shows a network that is hopelessly confused
by the information it has been given.
For a neural network, the learning process involves assigning weights to your input data. Information that
contributes to the network's ability to forecast correctly is assigned a corresponding weight based on its
value. Pairs of inputs and outputs are presented to the network. The network takes each pair and produces
an output, which it then compares with the correct forecast provided for training purposes. The
information that represents the correct forecast provided for training purposes is called a training pattern.