Backpropagation is the algorithm that calculates how each parameter in a neural network contributed to its error, so those parameters can be adjusted to improve. It is what makes training deep networks possible.
What it means in plain English
When a neural network makes a mistake, backpropagation works backward from the output through every layer, figuring out how much each parameter was responsible for the error. Those responsibility signals (gradients) then tell gradient descent how to adjust each parameter. The name comes from this backward flow of error information through the network.
Backpropagation and gradient descent work as a pair: backpropagation computes the direction to adjust; gradient descent takes the step.
A simple example
Think of a team that produces a flawed result. Backpropagation is like tracing back through each stage to assign a fair share of responsibility for the mistake — so each stage knows exactly how to do better next time.
Why it matters
Backpropagation is the algorithm that made training deep neural networks practical, and therefore underpins essentially all modern deep learning. Without an efficient way to assign blame across many layers, today’s large models could not be trained.
Related terms
- Gradient Descent — uses the gradients backpropagation computes.
- Neural Network — the structure backpropagation trains.
- Loss Function — defines the error being traced back.
Frequently asked questions
What does backpropagation actually do?
It calculates how much each weight in a neural network contributed to the error, so the weights can be adjusted to reduce that error — it is the core of how neural networks learn.
Is backpropagation the same as gradient descent?
No, but they work together: backpropagation computes the gradients (how the error changes with each weight), and gradient descent uses those gradients to update the weights.
