Idanwo ti o sopọ mọ itọsọna · Lile Ipele

Idanwo Ẹkọ Iyatọ

Practice contrastive positive pairs, embedding collapse, and temperature-scaled similarity, including how MoCo queues differ from inference caches.

Jẹmọ itọsọna awọn ọnaẸkọ Iyatọ
Ibeere 1 ti 8

Kini idi pataki ti ẹkọ iyatọ?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Kini idi pataki ti ẹkọ iyatọ?
    • Fa iru (rere) orisii papo ki o si Titari aiṣedeede (odi) orisii ni ifibọ aaye
    • Sọtẹlẹ awọn iye piksẹli gangan ti aworan kan
    • Sọtọ gbogbo igbewọle sinu ọkan ninu ṣeto awọn aami ti o wa titi
    • Din awọn nọmba ti awoṣe sile
  2. Ni ọna aworan ti o ni abojuto ti ara ẹni bii SimCLR, bawo ni a ṣe ṣẹda batapọ rere ni igbagbogbo?
    • Awọn aworan oriṣiriṣi meji ti kilasi aami kanna
    • Aworan ati aworan ti a yan laileto lati inu dataset
    • Awọn afikun laileto meji (fun apẹẹrẹ, awọn irugbin, jitter awọ) ti aworan kanna
    • Aworan ti a so pọ pẹlu iboju-ipinpin rẹ
  3. Iṣẹ ipadanu wo ni o ni nkan ṣe pẹlu ẹkọ iyatọ?
    • Aṣiṣe onigun mẹrin tumọ si
    • Alaye
    • Cross-entropy lori awọn akole kilasi
    • Hinge pipadanu fun SVMs
  4. Kini idi ti awọn ọna bii MoCo lo banki iranti tabi isinyi?
    • Lati tọju awọn akole fun iṣatunṣe didara to ni abojuto
    • Lati kaṣe awọn asọtẹlẹ ikẹhin fun itọkasi iyara
    • Lati compress awoṣe fun awọn ẹrọ alagbeka
    • Lati pese ọpọlọpọ awọn apẹẹrẹ odi laisi nilo iwọn titobi nla
  5. Iṣoro wo ni BYOL ati SimSiam ṣe aabo ni pataki laisi lilo awọn odi ti o fojuhan?
    • Overfitting si awọn aami ikẹkọ
    • Exploding gradients ninu awọn optimizer
    • Aṣoju ṣubu, nibiti gbogbo awọn ifibọ di aami
    • Oṣuwọn ẹkọ ti o padanu
  6. Bawo ni CLIP ṣe faagun ero iyatọ naa?
    • O ṣe deede awọn aworan pẹlu awọn akọle ọrọ wọn ni aaye ifibọ pinpin
    • O ṣe iyatọ awọn afikun meji ti agekuru ohun kanna
    • O yọ iwulo fun eyikeyi nẹtiwọọki nkankikan
    • O nlo ẹkọ imuduro dipo ibajọra
  7. A contrastive image encoder maps every photo to the same vector. Which failure does this illustrate?
    • A useful separation of image categories
    • An increase in the number of negative samples
    • A perfect nearest-neighbor representation
    • Representation collapse
  8. Why is MoCo’s training queue different from caching completed inference responses?
    • It serves final chatbot responses without model execution
    • It supplies encoded comparison examples for a contrastive training objective
    • It replaces every positive image with a label
    • It removes the need to update any encoder