Inside the Black Box: Understanding Neural Networks
Abstract
Artificial neural
networks (ANNs) have proven to be very valuable tools in such diverse
applications as robotic control, fraud detection, and spam
filtering. But-as valuable as ANNs are-they are also very
complex, and often poorly understood. They are generally
treated as "black box" systems-input is fed in and output is tested for
reliability without regard for how this transformation is
accomplished. However, a deeper understanding of exactly how
an ANN works can enhance its usefulness. This understanding is best
gained with a historical, hands-on approach, by examining the progress
of the methodology from the perceptron to the multilayer feedforward
network (MFN) with backpropagation by designing an example network and
putting it through its paces at each development level, as each step in
that progress has added a new layer to the networks, both of greater
generalization of usefulness and of greater complexity. This
deeper understanding of the inner workings of ANNs will demonstrate why
they have been so successful in certain types of applications, and what
sorts of applications they would likely be useful for in the future.

Diagram of MFN Successfully Trained to approximate XOR in Epoch 121

Diagram of MFN Successfully Trained to approximate XOR in Epoch 121
News :
2009-01-01
Happy New Year!!
2009-01-07
5th Draft Complete
2009-04-20
6th Draft Complete and submited
Contact Sergei Alderman