How do neural network connection weights change through experience?

In artificial neural networks, connection weights are adjusted through a mechanical process of training that allows the system to learn from experience without human intervention.1

Training typically involves comparing the network's actual output to a desired output. The weights are modified based on these rules:

• Increasing weights: If the actual output is 0 but the desired output is 1, the weights of connections from firing neurons are increased by a small, fixed amount. This makes the output neuron more likely to fire when the same pattern is presented again.1
• Decreasing weights: If the actual output is 1 but the desired output is 0, the weights are decreased by the same amount, making the output neuron less likely to fire for that pattern.1

More advanced systems, such as those using a back-propagation algorithm, can propagate these adjustments back through multiple layers of the network. Through repeated exposure to training examples, the network automatically forges a pattern of weights that enables it to respond correctly to inputs.21