Тест за хипотези за лотарийни билети

Отговорите и обясненията се появяват, докато отговаряте на теста.

1. Какво е „печеливш билет“ в тази хипотеза?
   • Разредена подмрежа, която, обучена от оригиналния си init, съответства на пълната мрежа
   • Мрежа с повече параметри от обикновено
   • Предварително обучен модел, изтеглен от интернет
   • Произволно познат набор от хиперпараметри
2. In the original lottery-ticket experiments, what was paired with the discovered sparse mask when retraining a ticket?
   • Fresh random values for every surviving weight
   • The surviving weights’ matched original initialization
   • A different architecture for each pruning round
   • The final trained weights with no resetting
3. Кой предложи хипотезата за лотарийния билет?
   • Goodfelow (2014)
   • Крижевски, Суцкевер, Хинтън (2012)
   • Франкъл и Карбин (2018)
   • Силвър и колеги (2016)
4. Which procedure did the original Lottery Ticket Hypothesis paper use to identify winning tickets?
   • Randomly remove a fixed half of all weights once
   • Iterative magnitude pruning with a reset to the original initialization
   • Add dropout after each pruning round
   • Increase the learning rate after every epoch
5. How should the idea that overparameterization helps optimization find a trainable sparse structure be described?
   • A conjecture or proposed explanation, not a result established by the original experiments
   • A theorem proven for every neural-network architecture
   • A guarantee that all weights can be removed without loss
   • A claim that pruning eliminates the need for training data
6. What does a strong lottery-ticket result claim in its theoretical setting?
   • Every randomly initialized network already solves any task
   • A sufficiently overparameterized randomly initialized network can contain a pruned subnetwork that approximates a target without further weight training
   • Pruning always preserves the trained model’s accuracy
   • A sparse network never needs data
7. What does iterative magnitude pruning remove at each pruning step?
   • The training examples with the lowest labels
   • The neurons with the most outgoing connections
   • Weights with the smallest magnitudes
   • All weights in the final layer
8. Which datasets were used in the original paper’s reported lottery-ticket experiments?
   • ImageNet and COCO only
   • LibriSpeech and Common Voice
   • SQuAD and GLUE
   • MNIST and CIFAR-10