Backpropagation

Algoritma untuk melatih neural network multi-layer dengan menghitung gradien loss terhadap setiap bobot menggunakan chain rule. Dipopulerkan Rumelhart-Hinton-Williams 1986.

Backprop = forward pass → compute loss → backward pass (gradients) → update weights. Fondasi training hampir semua neural network modern.

Also known as: backprop
Print

Backpropagation

Definisi

Backpropagation adalah algoritma untuk menghitung gradien loss function terhadap setiap bobot dalam neural network, menggunakan chain rule kalkulus.

Langkah

  1. Forward pass — hitung output layer-by-layer
  2. Compute loss — bandingkan dengan target
  3. Backward pass — hitung gradien dari output ke input
  4. Update weights — w ← w - η·∇L (gradient descent)

Sejarah

  • 1970 — Linnainmaa (master’s thesis)
  • 1986 — Rumelhart, Hinton, Williams mempopulerkan
  • 1989 — LeCun menggunakan untuk CNN (LeNet)
  • 2010+ — GPU + ReLU membuat deep network bisa di-training

Varian

  • SGD (Stochastic Gradient Descent)
  • Momentum, Nesterov
  • Adam (Kingma & Ba, 2014) — paling populer
  • AdamW — dengan weight decay

Connected to

Not yet written

The following pages are referenced but don't exist yet — they'd make good future additions.

  • /concepts/neural-network

References

  1. Wikipedia

Type at least 2 characters to search.

Press to navigate, to open, esc to close.