Dead End Street

A deep feedforward network using ReLU activations has several neurons whose outputs are stuck at zero for every training example, and their weights stop updating entirely. Which property of ReLU is most directly responsible for this 'dying ReLU' problem?

ReLU's output range is unbounded above, causing exploding activations that get clipped to zero downstream

ReLU's gradient is exactly zero for any input less than zero, so a neuron pushed into that region gets no gradient signal to recover

ReLU saturates and produces near-zero gradients for large positive inputs

ReLU is not zero-centered, which slows convergence but does not zero out gradients