Clicked Gallery

What is Backpropagation?

Highlighted from a real engineering doc. Explained by Clicked.

Used in a sentence

During the model training loop, the network used backpropagation to pass loss metrics backward and adjust node weights.

The reader highlighted one word in the docs. Clicked broke down the technical term “backpropagation” into plain English:

Explained in three depths

●○○

Overview

▾

Backpropagation is the algorithm that works out how much each of a neural network's millions of internal settings contributed to a mistake, by passing the error backwards through the layers, so every one of them can be adjusted by its share. It's the piece that makes training deep networks possible at all.

●●○

Detail

▾

Training runs in two passes. Forward: the input flows through the layers to produce a prediction, and the loss measures how wrong it was. Backward: using the chain rule from calculus, that error travels back through the network layer by layer, computing each weight's exact share of the blame, called its gradient, in a single sweep. That sweep is the breakthrough, because testing millions of weights one at a time would be hopelessly slow, while backpropagation prices all of them at once. The gradients then go to an optimizer, typically stochastic gradient descent, which nudges each weight against its blame. Forward, backward, nudge, repeated per batch millions of times: backprop computes the directions and SGD takes the steps.

●●●

Analogy

▾

Excel's Goal Seek, at scale. Goal Seek takes a formula, a target, and one input cell, then finds the right input by trying a value, measuring the miss, and adjusting. Backpropagation is the version with millions of input cells, computing the precise adjustment for every one of them from a single measurement of how wrong the output was.

AI explanations may contain errors · Not professional advice

Same facts, different vibe — Slang mode 😎

●○○

Overview

▾

Backprop is the blame audit. The model messes up, and the error walks backwards through every layer going "you contributed 3% of this, you 0.4%, you 0.001%," and then every single weight adjusts by exactly its share. Run that a few million times and it starts looking like intelligence. 😎

●●○

Detail

▾

Forward pass: the guess. Loss: how bad the guess was. Backward pass: chain-rule receipts for every weight in the network, computed in one sweep, and that sweep is the whole miracle, because checking millions of knobs one at a time would run past the heat death of the universe. Receipts go to the optimizer, which is SGD's job, every knob gets nudged against its blame, and the next batch starts. Backprop finds the directions, SGD takes the steps, and training is just this loop with a massive electricity bill. 😎

●●●

Analogy

▾

Your group project got a C, and instead of the usual vibes-based finger-pointing, imagine an exact printout: which sentence, written by whom, cost how many points. Everyone fixes precisely their part, and the next draft is better. Backprop is that printout, for every mistake, millions of contributors at a time.

AI explanations may contain errors · Not professional advice

Formal definition — The same term, explained the usual way

Backpropagation is an algorithm for computing the gradient of a loss function with respect to all parameters of a neural network via reverse-mode automatic differentiation. Following a forward pass that produces predictions and a scalar loss, the chain rule is applied layer by layer in reverse to obtain each parameter's partial derivative in a single backward traversal, with cost comparable to the forward computation. The resulting gradients parameterize weight updates performed by an optimizer such as stochastic gradient descent.

Want Clicked to explain terms like “backpropagation” directly in your browser — including on PDFs?

Add to Chrome — Free

50 free Explanations · No credit card required

More from the gallery

How to actually read a research paper

Skim smart, break the jargon wall, and keep what you learned — one session, done right.

What is Retrieval-Augmented Generation (RAG)?

Make the AI answer from your documents — retrieval first, generation second, receipts included.

What are Vector Embeddings?

Meaning turned into coordinates — how machines find what's similar without matching a single word.

What does an Idempotent API mean?

Why ten clicks charge you once — the property that makes retries safe.

What is Technical Debt?

Shortcuts and aging systems that bill the future — principal, interest, and repayment.

What is a Zero-Knowledge Proof (ZKP)?

Prove something is true without revealing why — the cryptography behind private verification.