I am solving a multi-label audio classification task with neural networks. The dataset is comprised of 10 classes, and the input data to the network are audio files where two of these classes are overlapped. The audio files are presented as a spectrogram to the network, which you can think of as a 1-channel image that looks like this:
So given the problem, the vector of true labels y will look like [0,0,1,0,0,0,1,0,0,0]
I use 3 convolutional layers, 1 FC layer and as output layer I use 10 logistic units. I calculate the cost as the mean of the ten cross entropy losses (one for each logistic unit).
My concern is, will the network tend to classify zeros rather than ones? Since in the training data there are more zeros than ones?
Should I add some term to the cost function to correct this issue? Or this is not an issue at all.