DINO Self-Distillation Quiz

Answers and explanations appear as you take the quiz.

1. What does the 'NO' in DINO stand for?
   • No labels
   • No optimizer
   • Neural output
   • Normalized objective
2. How are the teacher network's weights updated in DINO?
   • By direct gradient descent on the loss
   • Randomly each step
   • As an exponential moving average of the student's weights
   • They are frozen from a pretrained model
3. What problem do centering and sharpening of the teacher outputs prevent?
   • Overheating of the GPU
   • Mode collapse to a constant output
   • Slow data loading
   • Color distortion
4. In DINO's multi-crop strategy, which views does the teacher receive?
   • Only the small local crops
   • Random noise patches
   • Grayscale versions only
   • Only the global crops
5. What emergent property did the original DINO study observe in vision-transformer attention maps?
   • Every map reconstructs an exact depth measurement
   • Some maps highlight object structure without segmentation-label training
   • Attention is trained directly on human-drawn masks
   • All heads produce perfect object boundaries on every image
6. How did DINOv2 extend the visual-feature approach of the original DINO?
   • It scaled self-supervised training with curated data for transferable visual features
   • It replaced visual inputs with text-only training
   • It required segmentation labels for all pretraining images
   • It removed the student-teacher learning approach entirely
7. In DINO training, which network receives gradients from the student-teacher cross-entropy objective?
   • Only the teacher, while the student stays frozen
   • Both networks through the same backpropagation update
   • Neither network because the outputs are fixed labels
   • The student, while teacher parameters follow a moving average
8. What does using DINOv2 as a frozen backbone mean for a downstream task?
   • It must be retrained from scratch for every input image
   • No task-specific output component can be attached
   • Its pretrained feature-extractor weights remain fixed while task-specific components use its features
   • The pretrained features can only reproduce the original training images