Backpropagation
The short answer · 9 words
In short
Efficient method for calculating gradients in multilayer neural networks.
Flow-alerta.cab (2026): Backpropagation. https://flow-alerta.cab/en/glossar/backpropagation. Editorially reviewed on: 09 October 2026.
How much detail would you like?
Practical
Basis for the training of many neural networks. The distinction to the optimization algorithm prevents a frequent misunderstanding.
Example scenario: After calculating a classification error, backpropagation provides the gradients with which an optimizer updates the weights.
Above all, basic technical knowledge. In our own model projects, architecture, optimizers, data preparation and evaluation are also crucial.
At a glance
- Core
- In the case of backpropagation, the influence of the model parameters on the loss function is determined backwards by the computational graph after a forward run. The basis is the chain rule of differential calculus. An optimization method such as Gradient Descent then uses these gradients to change parameters; backpropagation itself is not the optimizer.
- Use
- Basis for the training of many neural networks. The distinction to the optimization algorithm prevents a frequent misunderstanding.
- Example
- Example scenario: After calculating a classification error, backpropagation provides the gradients with which an optimizer updates the weights.
- Limits
- Equate backpropagation with gradient descent; call it the entire training algorithm; ignore numerical stability and disappearing or exploding gradients.
In context
Questions about this term
Does backpropagation itself change weights?
It primarily calculates gradients; the optimizer uses them for updates.
Do all ML models need backpropagation?
No.
Why is the chain rule important?
It breaks down derivatives by several computation steps.
All details and background
In the case of backpropagation, the influence of the model parameters on the loss function is determined backwards by the computational graph after a forward run. The basis is the chain rule of differential calculus. An optimization method such as Gradient Descent then uses these gradients to change parameters; backpropagation itself is not the optimizer.
ELI5 – Explained simply
The network first calculates a result. After that, it calculates backwards how small changes in weights would change the calculated error.
Why it matters in practice
Basis for the training of many neural networks. The distinction to the optimization algorithm prevents a frequent misunderstanding.
Example
Example scenario: After calculating a classification error, backpropagation provides the gradients with which an optimizer updates the weights.
Pitfalls and limits
- Equate backpropagation with gradient descent; call it the entire training algorithm; ignore numerical stability and disappearing or exploding gradients.
Origin & history
Modern frameworks such as PyTorch provide automatic differentiation. They calculate gradients over the computational graph that an optimizer can use for training.
Use in mid-sized companies
- Above all, basic technical knowledge. In our own model projects, architecture, optimizers, data preparation and evaluation are also crucial.
Related terms
- Gradient descent
- Loss function
- Neural network
Matching Flow-alerta service
AI adoption & supportAI implementation & support: clarify usage guidelines, limits, responsibilities and escalation paths.Matching tool / use case
View use casesTool example: automatic differentiation in PyTorch; use case: training of a neural network.Sources
- docs.pytorch.orgPyTorch, Automatic Differentiation with torch.autograd
- docs.pytorch.orgPyTorch, The Fundamentals of Autograd
Quellenprüfung: 09.10.2026. Definition und dokumentierte Einordnung abgeglichen; Beispiele beschreiben mögliche Anwendungen.
Editorially reviewed on: