Maswali yanayounganishwa na mwongozo · Rahisi Kiwango

Maswali ya Kusimamisha Mapema

Inashughulikia ufuatiliaji wa uthibitishaji, subira na uokoaji wa kituo cha ukaguzi, na wakati kuacha mapema kunaweza kuhalalisha mafunzo bila kuhakikisha makosa ya chini ya jaribio.

Njia za mwongozo zinazohusianaKuacha Mapema
Swali 1 ya 8

Ni ishara gani kuacha mapema hufuatilia hasa kuamua wakati wa kusitisha?

Hakiki kila swali8 maswali

Majibu na maelezo huonekana unapojibu maswali.

  1. Ni ishara gani kuacha mapema hufuatilia hasa kuamua wakati wa kusitisha?
    • Utendaji uliowekwa wa uthibitishaji
    • Upotezaji wa seti ya mafunzo
    • Halijoto ya GPU
    • Idadi ya vigezo vya mfano
  2. Je, parameta ya 'uvumilivu' inadhibiti nini?
    • Kiwango cha juu cha kujifunza
    • Ni nyakati ngapi za kusubiri bila uboreshaji kabla ya kusimama
    • Ukubwa wa kundi
    • Idadi ya mikunjo ya uthibitishaji
  3. Baada ya vichochezi vya kuacha mapema, ni uzito gani unapaswa kuweka kwa kawaida?
    • Vipimo kutoka enzi ya mwisho
    • Vipimo vilivyoanzishwa kwa nasibu
    • Vipimo kutoka enzi bora ya uthibitishaji
    • Wastani wa enzi zote
  4. Why can early stopping act as a form of regularization?
    • It adds a fixed L2 penalty to every weight
    • It guarantees lower test error on every dataset
    • It prevents the model from seeing training examples
    • It limits how long optimization proceeds and can reduce fitting to training-specific patterns
  5. Umbo la kawaida la curve ya uthibitishaji-kosa wakati wa mafunzo linafafanuliwa vyema kama:
    • U-umbo ambalo huanguka kisha huinuka
    • Kupungua kwa monotonically
    • Mstari wa gorofa
    • Kuongezeka kwa monotonically
  6. Je, kiwango cha juu cha 'min_delta' kwa kawaida kinabainisha nini?
    • Idadi ya chini ya tabaka
    • Kiwango cha chini cha uboreshaji ambacho kinahesabiwa kama maendeleo
    • Ukubwa wa chini wa kundi
    • Kiwango cha chini cha ukubwa wa seti ya data
  7. What does patience allow an early-stopping rule to do?
    • Continue training forever regardless of validation results
    • Change the model architecture after every batch
    • Wait through a set number of validation checks without qualifying improvement before stopping
    • Use the test set to select the final epoch
  8. Why should the test set generally remain outside the early-stopping loop?
    • The test set cannot contain labels
    • Training libraries cannot load test data
    • Early stopping requires the test set to be empty
    • Repeatedly selecting by test performance would make it part of model selection