What is the name of the procedure that allows pre-trained weights to be reused for the target task? What are the new hyperparameters that this approach requires?

Draw a simple Recurrent Network, with input x, outputy and recurrent state_h . Update the equations for a simple RNN unit in terms ofx,y, h . Use tanh non-linearity.?

Is an autoencoder necessarily a good autoencoder if it fully reconstructs the inputs? How can you evaluate an autoencoder's performance?