COURSE 05L1100% FREE
Verified 2026-08-10

Backpropagation

**Backpropagation** is how a neural net figures out **which weights to tweak** after a mistake. Errors flow backward through the layers so each weight gets a “this way / that way” signal.

What it is

Backpropagation is how a neural net figures out which weights to tweak after a mistake. Errors flow backward through the layers so each weight gets a “this way / that way” signal.

Why it matters

Without an efficient way to compute those signals, deep nets would be impractical to train. Backprop + gradient descent is the workhorse of deep learning.

How it works (plain)

  1. Forward pass: make a prediction
  2. Compare to the target → loss
  3. Backward pass: compute how tiny weight changes would change the loss
  4. Step weights to reduce loss
  5. Repeat

You do not hand-tune millions of weights; calculus + chain rule does the bookkeeping.

Everyday example

Like adjusting each knob on a giant soundboard after hearing the mix is too bass-heavy—except the “ears” are the loss function.

Try it

Explain to a friend: “Forward to guess, backward to assign blame to each knob, nudge knobs, repeat.”

Myths

⚠️ Myth: Backprop is a mysterious AI invention from 2020.
✓ Reality: The method is decades old; scale made it newly decisive.
⚠️ Myth: Backprop “understands” why the answer is wrong.
✓ Reality: It propagates gradients for a chosen loss—not human meaning.

Sources