If the deep learning neural network has the problem of gradient disappearance or gradient explosion, our common solution is.

Question

Options

• AGradient shear
• BRandom undersampling
• Cuse Relu Activation function
• DRegularization

• A
  82% (18)
• B
  18% (4)

Explanation

Gradient vanishing occurs when gradients become too small to update early layers, while gradient explosion causes unstable, runaway weight updates - both cripple training in deep networks. Gradient clipping (A) directly tackles explosion by capping gradient magnitudes before the backward pass, preventing parameter updates from spiraling out of control. ReLU (C) solves vanishing gradients by having a constant derivative of 1 for positive inputs, unlike sigmoid/tanh whose small derivatives compound across layers and shrink to near-zero. Regularization (D) - particularly L2 weight decay and batch normalization - constrains weight magnitudes and normalizes activations, stabilizing gradient flow throughout training.

B (Random undersampling) is wrong because it's a data preprocessing technique for fixing class imbalance, which has nothing to do with how gradients propagate through a network.

Memory tip: Think "Clip, Activate, Constrain" - you clip runaway gradients (A), activate with a non-saturating function (C), and constrain weights to stay well-behaved (D). All three attack the same root cause: gradients either vanishing or exploding during backpropagation.

No community discussion yet for this question.