Stable Deep-Network Training

Five questions. Choose one answer for each, then submit to see your score and explanations.

The practice target is 80%. You can retry as often as you like, and every lesson stays accessible.

You can practice as a guest. Sign in before starting an attempt to save your scores.

1. Why avoid initializing all hidden units with identical weights?A. Identical weights always cause a syntax errorB. All initial weights must be positiveC. Initialization determines the labels automaticallyD. Identical units can remain symmetric and learn the same features

2. Which initialization is commonly suited to ReLU networks?A. Every weight set to the same large constantB. Copying test labels into weightsC. He initializationD. All weights set to one

3. How does ordinary batch normalization typically handle inference?A. It drops all normalized activationsB. It uses stored running statistics and learned scale and shiftC. It always recomputes training gradientsD. It must use the final test labels

4. Gradients become extremely large and updates become unstable. Which technique directly limits gradient magnitude?A. Gradient clippingB. Removing the validation setC. Increasing every weight by a fixed constantD. Applying stronger image compression only

5. A ReLU unit receives only negative pre-activation values and outputs zero. What training issue can result?A. It necessarily outputs probabilities summing to oneB. It always has exploding positive activationsC. It automatically becomes a sigmoid unitD. It can receive zero activation gradients and stop contributing useful updates

Review this module from its first lesson →

Stable Deep-Network Training

Your result