Barlow Twins dan Kuis Pengurangan Redundansi

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. What does Barlow Twins compare during representation learning?
   • Embeddings from two augmented views of the same image
   • A labeled image and a text caption
   • A positive view against many negative examples
   • Two outputs from unrelated tasks
2. What target does the Barlow Twins objective push the cross-correlation matrix toward?
   • A matrix of all zeros
   • The identity matrix
   • A matrix of all ones
   • A diagonal matrix with random entries
3. What does a diagonal term in the objective encourage?
   • Different labels for augmented views
   • Independence from the other view
   • Similar features across the two views of the same example
   • A larger number of negative pairs
4. What does the off-diagonal penalty encourage?
   • Higher similarity between every feature pair
   • More negative image examples
   • A batch-dependent class label
   • Less redundancy among different embedding components
5. How does Barlow Twins differ from methods that require negative pairs?
   • It does not need negative examples in its objective
   • It uses more negative examples per image
   • It replaces negatives with class labels
   • It needs a separate model for every class
6. Which mechanism does Barlow Twins use to avoid constant collapsed representations?
   • An explicit set of negative labels
   • The redundancy-reduction objective on the cross-correlation matrix
   • A frozen target encoder updated by EMA
   • A supervised classification head
7. Does Barlow Twins require a predictor and stop-gradient asymmetry like BYOL?
   • Yes, both are required in the original method
   • Only when the batch contains one image
   • No; its paired networks use the correlation objective without that asymmetry
   • It uses a discriminator instead
8. What is a careful claim about Barlow Twins benchmark performance?
   • It will outperform every supervised method on all tasks
   • It proves correlation objectives are always best
   • It guarantees useful features without downstream testing
   • It performed well in the specific image evaluations reported by its authors