Backpropagation Weight Training Rule

• the error in the delta rule is replaced by
• for output unit it is the familiar from the delta rule multiplied by which is derivative of the sigmoid squashing function
• for hidden unit the derivative component is the same but there is no target value directly available so you sum the error terms for each output unit influenced by weighting each of the by the weight, , from the hidden unit to the output unit .
• This weight characterizes the degree to which each hidden unit is responsible for the error in output unit .