Given that neural networks are to support nonlinearity and more complexity, the activation function to be used has to be robust enough to have the following:
- It should be differential; we will see why we need differentiation in backpropagation. It should not cause gradients to vanish.
- It should be simple and fast in processing.
- It should not be zero centered.
The sigmoid is the most used activation function, but it suffers from the following setbacks:
- Since it uses logistic model, the computations are time consuming and complex
- It cause gradients to vanish and no signals pass through the neurons at some point of time
- It is slow in convergence
- It is not zero centered
These drawbacks are solved by ReLU. ReLU is simple and is faster to process. It does not have the vanishing gradient problem and has shown vast improvements compared to the sigmoid and tanh functions. ReLU is the most preferred activation function for neural networks and DL problems.
ReLU is used for hidden layers, while the output layer can use a softmax function for logistic problems and a linear function of regression problems.
