Neural networks are based on a mathematical technique that is inspired by the parallel processing power of the human brain. Unlike other techniques, neural nets have the remarkable ability to learn. As such, they can be trained to recognize patterns and detect trends in the data that cannot be discerned by conventional means.


Neural Network: Analysis
The neural net first has to be constructed, taylored to the problem at hand. Several parameters need to be chosen. These are, for example, the type of input data; the elements or patterns to train for; the number of nodes, and the number of layers. The top panel above shows the lay-out of each node as well as the node layers in our example of pattern recognition. The inputs Xi are weighted during the training phase, and the perceptron output is filtered with a pattern function H(U).


Neural Network: Results
Here, the neural net has been trained to recognize so-called shocklets in the data. The shocklets (see arrows) are embedded in random noise, and have two distinguishing features: (a) a sharp edge on the left side, and (b) high frequency waves in the vicinity of this front. A value of "1" in the lower panel indicates that the neural net proclaimed the corresponding part of the data to contain a shocklet. Success! ...The neural net picked the correct regions.