Is it possible to use discontinuous activation functions in a neural network?

If an activation function has a jump discontinuity, then in the training process, can we implement backpropagation to compute the derivatives and update the parameters?

10 January 2024 7,209 1 View