Ikibazo gihuza ibibazo · Hagati Urwego

Stochastic Gradient Kumanuka hamwe na Momentum Ikibazo

Ikizamini cyegeranije umuvuduko, kugabanuka kunyeganyega, ingaruka za coefficient, gradients zihoraho, guhuza Adamu, nimyitwarire ya zeru.

Ikibazo 1 Bya 8

Niki ijambo ryihuta ryegeranya mugihe cy'amahugurwa?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Niki ijambo ryihuta ryegeranya mugihe cy'amahugurwa?
    • Umubare rusange wibihe byamahugurwa
    • Ikigereranyo (kiremereye cyane) impuzandengo ya gradients zashize
    • Ingano yamakuru
    • Icyitegererezo cyamateka
  2. Mu kibaya kirekire, kigufi, ni ikihe kibazo SGD isanzwe ihura nizo mbaraga zifasha gukemura?
    • Yuzuye mu mubare
    • Irahagarika kubara gradients
    • Yongera igipimo cyo kwiga mu buryo bwikora
    • Zig-zags hejuru yinkuta zihanamye mugihe igenda gahoro gahoro hasi
  3. Agaciro gasanzwe kubiciro bya coefficient ni:
    • 0.9
    • 0.0
    • 0.5
    • 2.0
  4. Ni ikihe kigereranyo gifatika gikoreshwa cyane mugusobanura imbaraga?
    • Pendulum izunguruka ubuziraherezo
    • Isoko isubira inyuma
    • Umupira uremereye umanuka
    • Gukonjesha amazi
  5. Nkanswe ni kangahe imbaraga zongerera intambwe ifatika icyerekezo gihamye mugihe beta = 0.9?
    • Inshuro zigera kuri 2
    • Inshuro zigera kuri 0.1
    • Nta amplification
    • Inshuro zigera ku 10
  6. Ni ubuhe buryo bugezweho bukubiyemo imbaraga-yuburyo bwa mbere?
    • Uburyo bwa Newton
    • Adam
    • K-bisobanura
    • Gushakisha bisanzwe
  7. In the update v = beta times v plus gradient, what happens when beta is zero?
    • The old velocity is doubled
    • Every gradient becomes zero
    • The previous velocity is discarded and the current gradient determines the step
    • The parameters stop depending on the learning rate
  8. Why can repeated gradients with opposing sideways directions produce less zig-zagging under momentum?
    • Their opposing contributions can partially cancel in accumulated velocity
    • All past gradients are stored as separate parameters
    • The loss becomes convex automatically
    • Momentum ignores the current gradient entirely