#dying-relu

Question 1: Single choice

DCGAN's discriminator uses leaky ReLU, with slope $\alpha = 0.2$ for negative inputs. What is the main risk of using a standard ReLU there instead?

The discriminator can no longer output probabilities, because ReLU outputs have no upper bound1
The discriminator becomes a linear function of its input, so it cannot separate real from fake2
Units whose inputs stay negative output 0 with zero gradient, stop learning, and pass the generator less signal3
Negative pixel values of the generated images are clipped to 0 before the discriminator sees them4