Ikibazo gihuza ibibazo · Hagati Urwego

Ikizamini cya Gradient

Kugenzura ibisanzwe no kugabanya agaciro, kubika icyerekezo, kugenzura-gutondeka-kugenzura no kurugero rwo gukuramo mumahugurwa atandukanye.

Inzira zijyanye no kuyoboraGucisha make
Ikibazo 1 Bya 8

Ni ubuhe buryo clip-by-isanzwe igabanya gukuramo cyane cyane kubika mugihe bigabanya ivugurura?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Ni ubuhe buryo clip-by-isanzwe igabanya gukuramo cyane cyane kubika mugihe bigabanya ivugurura?
    • Icyerekezo cya rusange igenda ivugururwa
    • Ubunini nyabwo bwa buri cyiciro
    • Umubare w'icyitegererezo
    • Gahunda yo kwiga
  2. Muri clip-by-norm hamwe na bipimo c, bigenda bite mugihe igipimo gisanzwe g_norm kirenze c?
    • Gradients yashyizwe kuri zeru
    • Buri cyiciro kigwizwa na c / g_norm
    • Amahugurwa arahagarara
    • Igipimo cyo kwiga cyikubye kabiri
  3. Ni ikihe kibazo gukata buhoro buhoro bigenewe kurwanya?
    • Kurimbuka
    • Buhoro buhoro gupakira amakuru
    • Ibikoresho biturika
    • Kurenza urugero rwo kwemeza
  4. Nigute clip-by-agaciro itandukaniye na clip-by-bisanzwe?
    • Burigihe burinda icyerekezo cyo kuvugurura
    • Ikora gusa kumurongo wemeza
    • Ikuraho gradients burundu
    • Ifunga buri cyiciro kigizwe n'ubwigenge mu ntera, ishobora guhindura icyerekezo
  5. A minibatch produces an unusually large global gradient norm. What does clip-by-norm do before the optimizer update?
    • Scales the full gradient vector down to the configured threshold when the norm exceeds it
    • Changes the model’s labels
    • Drops every gradient component independently
    • Increases the learning rate until the loss recovers
  6. Mu mahugurwa atandukanye (DP-SGD), ni uruhe ruhare kurugero rwo gukuramo buhoro buhoro?
    • Irabika amakuru
    • Ihuza urugero rwa buri rugero kuburyo urusaku rushobora kwemeza ubuzima bwite
    • Yongera ingano yicyiciro
    • Bikuraho gukenera imikorere yigihombo
  7. The global gradient norm is twice a clipping threshold c. By what factor is the gradient scaled?
    • Babiri
    • One quarter
    • One half (c divided by 2c)
    • It is left unchanged
  8. Why does DP-SGD clip each example’s gradient before adding calibrated noise?
    • To make all examples have the same label
    • To remove the need for a privacy budget
    • To increase the norm of rare-class gradients
    • To bound an individual example’s influence on the released update