6.034 Artificial Intelligence - Recitations, fall 2004 online slides on learning
	Learning in a Multilayer Neural Network How should we change (adapt) the weights in a multilayer neural network? A one layer net is easy - there is a direct mapping between weights and output. Here, weights can contribute to intermediary functions and only indirectly affect output. These weights can indirectly affect multiple output nodes. If output is 12 and we want a 10, change weights to output node so that output next time would be (closer to) desired value 10. How do we change weights to hidden units? Assign portion of error to each hidden node, change weights to lessen that error next time.