Training with Back Propagation

Abstract

What does it mean to “train” a neural network? The network model we are using “learns” or gets better at its task by adjusting its connection strengths. We will strengthen those signals that tend to contribute to a right answer, and weaken those signals that tend toward a wrong answer. We do this by adjusting the potentiometer that lies in the path of each signal. It’s kind of like changing the conductive environment in a neuron’s synapse. To accomplish this, we will simply increase or decrease the voltage in question by 0.2V. That’s our method of implementing back propagation of errors. Pretty crude, I know, but it makes for a good simulation of the process.