The argmax function returns the index of the maximum value in the input vector. The output of argmax looks like this:
Argmax is used to narrow down the final classification for various activation functions like the Softmax Activation Function.
Limitations
One limitation of the argmax function is that all its gradients are always zero. Since backpropagation is dependent on calculating a loss function based on the misclassified elements in the vector, argmax cannot be used during the training of a neural network as there will be no learning.