Maswali yanayounganishwa na mwongozo · Ngumu Kiwango

Maswali ya BYOL na Maswali ya Kujisimamia Isiyo ya Kinyume

Hujaribu matawi ya mtandaoni na lengwa ya BYOL, mionekano iliyoboreshwa, utabiri, masasisho ya wastani yanayosonga, na mafunzo ya uwakilishi bila lebo.

Njia za mwongozo zinazohusianaByol na Usimamizi wa kibinafsi usio na Kinyume
Swali 1 ya 8

What makes BYOL non-contrastive?

Hakiki kila swali8 maswali

Majibu na maelezo huonekana unapojibu maswali.

  1. What makes BYOL non-contrastive?
    • It trains only with class labels
    • It learns without using negative image pairs
    • It does not use image augmentations
    • It compares each image with every other image
  2. What are BYOL’s two networks called?
    • Online and target networks
    • Encoder and classifier only
    • Teacher and discriminator
    • Query and key databases
  3. What does the online network learn to predict?
    • The image’s human class label
    • A negative pair’s representation
    • The target network’s raw pixel values only
    • The target network’s representation of another augmented view of the same image
  4. How are target-network weights updated?
    • By ordinary gradient descent on the same loss
    • They remain random and frozen forever
    • By a slow-moving exponential average of online weights
    • By copying the classifier head each batch
  5. Why does BYOL use an online predictor and stop-gradient on the target side?
    • They create an asymmetric learning setup that helps avoid representation collapse
    • They add negative examples
    • They turn the task into supervised classification
    • They make the target receive the same gradient
  6. What happens to two views of the same image in BYOL?
    • They are assigned different class labels
    • They are encoded in the two networks and used to align learned representations
    • They are compared only to unrelated images
    • They are discarded after augmentation
  7. Does BYOL require human class labels during its representation pretraining?
    • Yes, every view needs a class label
    • Only the target branch is labeled
    • No; it is a self-supervised image-representation method
    • It needs labels for negative pairs
  8. What is a fair interpretation of BYOL’s benchmark results?
    • They prove labels are never useful
    • They guarantee superiority on all datasets
    • They show that all non-contrastive methods work identically
    • They demonstrate performance on specified evaluations, not a guarantee for every task