Ikibazo gihuza ibibazo · Hagati Urwego

Ikibazo cya Backpropagation

Gerageza gusubiza inyuma gradients, guhinduranya kwishingikirizaho, gutondekanya imbere indangagaciro, urunigi-amategeko kugwiza, gusangira-inzira amafaranga, hamwe no gutandukanya optimizer.

Inzira zijyanye no kuyoboraGusubira inyuma
Ikibazo 1 Bya 8

What quantity does backpropagation compute for a neural network's relevant weights?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. What quantity does backpropagation compute for a neural network's relevant weights?
    • Derivatives of the loss with respect to those weights
    • A causal percentage of total error uniquely assigned to each weight
    • The architecture drawing without any calculation
    • A permanently updated set of weights without an optimizer
  2. Ni irihe tegeko ry'imibare riri mu mutima wo gusubira inyuma?
    • Inyigisho ya Pythagorean
    • Bayes 'theorem
    • Amategeko y'urunigi rwo kubara
    • Inzira ya kane
  3. Why can reverse-mode backpropagation compute many parameter derivatives efficiently?
    • It avoids all local derivatives
    • It reuses graph structure and relevant intermediate results
    • It calculates only the output prediction again
    • It guarantees every implementation has exactly one forward pass of runtime
  4. Ni mu buhe buryo gusubira inyuma bibara gradients?
    • Kuva mubyinjijwe byerekeza kubisohoka
    • Inzego zose icyarimwe nta gahunda
    • Gusa kuri hagati rwihishe
    • Uhereye kubisohoka bisubira inyuma ugana ibyinjira
  5. Does computing backpropagation gradients by itself establish that model weights were updated?
    • Yes; every backward computation automatically applies a learning rate
    • Yes; gradient values always replace the weights directly
    • No; the optimizer update is a separate step
    • Only if the network has no parameters
  6. In a conventional cached-backpropagation workflow, what does the forward computation provide for later derivative calculations?
    • Relevant intermediate values that can be retained for the backward computation
    • A completed optimizer update without any update rule
    • A guarantee that all later evaluation errors disappear
    • A permanent deletion of the training objective
  7. On one composed backpropagation path, local scalar derivatives are three and four. What is their chain-rule product?
    • Birindwi
    • Imwe
    • Three quarters
    • Twelve
  8. A shared parameter has loss-gradient contributions two and five through two paths. What combined contribution is accumulated?
    • Icumi
    • Birindwi
    • Bitatu
    • Two fifths