All-zero hidden-layer initialisation

Why is assigning identical zero weights to every hidden unit generally a poor neural-network initialisation?

Symmetric units can receive identical gradients and remain identical, preventing the layer from learning distinct representations. Depending on activations and architecture, zero derivatives can also block learning. Use an appropriate asymmetric variance-scaled initialisation with a recorded RNG seed. Zero biases are often acceptable; the symmetry problem concerns equivalent hidden units, not a universal ban on every zero parameter.