Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele

Idanwo Amugbamu-ašẹ

Test source and target shifts, reversed domain gradients, target-label scarcity, source-free methods, and separate classifier objectives.

Jẹmọ itọsọna awọn ọnaImudara-ašẹ
Ibeere 1 ti 8

Iṣoro wo ni aṣamubadọgba agbegbe ni akọkọ ṣe apẹrẹ lati yanju?

Awotẹlẹ gbogbo ibeere8 ibeere

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

  1. Iṣoro wo ni aṣamubadọgba agbegbe ni akọkọ ṣe apẹrẹ lati yanju?
    • Atehinwa awọn nọmba ti paramita awoṣe
    • Iyara isodipupo matrix lori awọn GPU
    • Ilọkuro iṣẹ ṣiṣe nigbati data idanwo yato si data ikẹkọ
    • Ṣiṣẹda awọn aami ikẹkọ sintetiki laifọwọyi
  2. Ninu nẹtiwọọki atako-ašẹ kan, kini ipele ipadasẹhin gradient ṣe aṣeyọri?
    • O titari ẹya ara ẹrọ jade lati ṣe orisun ati ibi-afẹde ko ṣe iyatọ
    • O npa olupilẹṣẹ agbegbe kuro lakoko itọkasi
    • O ṣe ilọpo meji oṣuwọn ikẹkọ fun ikẹkọ yiyara
    • O ṣe iyipada awọn abajade ifasilẹyin sinu awọn aami ikasi
  3. Kini o jẹ ki aṣamubadọgba agbegbe 'alabojuto' wulo paapaa?
    • Ko nilo ikẹkọ awoṣe rara
    • O ṣiṣẹ nikan lori data aworan
    • O ṣe imukuro iwulo fun aaye orisun kan
    • Ko nilo data isamisi ni agbegbe ibi-afẹde
  4. Oju iṣẹlẹ wo ni apẹẹrẹ Ayebaye ti iṣipopada agbegbe?
    • Awoṣe ikẹkọ ati idanwo lori dataset gangan kanna
    • Atọka ikasi tumo ti oṣiṣẹ lori ẹrọ ọlọjẹ kan kuna lori ẹrọ iwoye ti o yatọ
    • Awoṣe ti o gbalaye yiyara lẹhin quantization
    • Nẹtiwọọki pẹlu awọn ipele diẹ sii ju nẹtiwọki miiran lọ
  5. Kini ọna imudọgba agbegbe 'ọfẹ-orisun' ti a ṣe apẹrẹ lati yago fun?
    • Iwọle si data ikẹkọ orisun atilẹba lakoko aṣamubadọgba
    • Lilo eyikeyi faaji nẹtiwọki nkankikan
    • Ṣiṣejade awọn asọtẹlẹ iṣejade eyikeyi
    • Nṣiṣẹ lori igbalode GPUs
  6. A team starts adaptation from a broadly pretrained model. What should it still measure?
    • Only the parameter count
    • Only source training accuracy
    • Performance on the actual target data and task
    • Whether target evaluation can be skipped
  7. Which DANN head supplies the objective to distinguish source examples from target examples?
    • The task label vocabulary
    • The domain classifier
    • The optimizer checkpoint file
    • A fixed output embedding dimension
  8. Does gradient reversal mean changing a positive class label into a negative class label?
    • Yes; every task label is negated
    • Yes; all target samples are relabeled as source
    • No; it freezes every feature weight
    • No; it reverses the domain gradient into the feature extractor