Model Pruning Quiz

Answers and explanations appear as you take the quiz.

1. What does model pruning remove from a trained network?
   • Removing weights or structures that contribute little to the output
   • Adding more parameters to improve accuracy
   • Training a model for more epochs
   • Increasing the input resolution
2. What distinguishes structured pruning from unstructured pruning?
   • Structured pruning only works on images
   • Structured pruning removes whole units like neurons or heads; unstructured removes individual weights
   • Unstructured pruning requires retraining from scratch
   • There is no practical difference between them
3. Why does unstructured pruning often fail to speed up a model on ordinary hardware?
   • It deletes too many layers
   • It always reduces accuracy to zero
   • Sparse weight matrices need special hardware or libraries to exploit the zeros
   • It increases the parameter count
4. What is the typical iterative pruning recipe?
   • Prune everything at once, then deploy
   • Quantize first, then never retrain
   • Randomly delete layers until accuracy drops
   • Train, prune least-important parameters, fine-tune to recover, and repeat
5. What does the Lottery Ticket Hypothesis claim?
   • Dense networks contain sparse subnetworks that can match the full model when trained from the right initialization
   • Pruning always improves accuracy
   • Random pruning is always best
   • Large models cannot be compressed
6. Which simple criterion is commonly used to decide which weights to prune?
   • The alphabetical order of layer names
   • The absolute magnitude of the weights
   • The color of the activation maps
   • The number of training examples
7. When is pruning most likely to reduce inference latency on ordinary dense hardware?
   • When individual zeros are added but the dense kernel still computes them
   • When the model is renamed without changing its graph
   • When whole channels or other structures are removed to produce a smaller dense model
   • When a sparse checkpoint is stored in a larger file
8. What should a team measure after pruning for a deployment target?
   • Only the number of zeros in a checkpoint
   • Only training accuracy before pruning
   • Whether the file extension changed
   • Task quality and latency or memory on the intended hardware