measure of the error contributed by each neuron in neural network

the NN is evaluated for a random training instance using forward propagation, and the error at the network output (with reference to the golden output) is computed. For each instance in the training dataset, the error at the output of the neural network is computed using forward propagation. Next, backpropagation is invoked, the errors are propagated back to the outputs of individual neurons and their average error contribution over all inputs in the training set is obtained. Thus, backpropagation quantifies the error contributed by each neuron in the network towards the global network error.

how can I provide a measure of the error contributed by each neuron to the outputs of the network by backpropagation in matlab? I would appreciate for any help.

Answer 1

[Greg Heath]

1. Simple approach:

For each hidden node:

 Zero all input weights to the singly selected hidden node and record the performance.
 Rank the performances

2. Less simple approach.

 Use the simple approach to rank existing nodes
 Remove the node yielding the worst performance
 Repeat until one node is left

3. Complicated approach:

 Repeat 2 EXCEPT: Retrain the network after each node is removed.

Hope this helps.

Greg