I am using a neural network with two output neurons for my omics data to predict the binary classification of genes. It is important in the study to know the feature importance. I know that the Garson's algorithm and connection weights algorithm are used in neural networks with one hidden layer, but in all the examples I've seen there is only one output neuron. My neural network has two output neurons. Can I still use these algorithms to evaluate feature importance in my neural network?
I could be wrong but I sense a misunderstanding here. A neural network for binary classification typically learns an output distribution, not a function. So their should only be one output, probability of a positive output given the inputs. 1 minus this value would then be the probability of a negative output. It is done this way since gradient descent with back propagation needs smooth outputs.
Multiple outcomes are useful if the classification is not binary. Perhaps there are two dimensions being evaluated simultaneously. Multiple classification can use as many outputs as classes in a "one-hot" encoding. If the training algorithm needs a single value for goodness of fit, the multiple output neurons should be summarized.
I agree. But in this case he said 'binary classification'.