Cross-Entropy Loss
The short answer · 13 words
In short
Loss function that compares predicted class distributions to target classes or target distributions.
Flow-alerta.cab (2026): Cross-Entropy Loss. https://flow-alerta.cab/en/glossar/cross-entropy-loss. Editorially reviewed on: 09 October 2026.
How much detail would you like?
Practical
Relevant for training classifiers and predicting tokens.
Example scenario: A ticket model learns concern classes based on labeled examples with cross-entropy loss.
Ensure label definition and data quality; additionally evaluate confusion matrix, recall and technical error costs on independent data.
At a glance
- Core
- For a unique target class, cross entropy penalises a low predicted probability for that class. For multiclass classification, PyTorch CrossEntropyLoss expects unnormalized logits and processes the necessary normalization internally. Training loss and practical classification quality are different sizes.
- Use
- Relevant for training classifiers and predicting tokens.
- Example
- Example scenario: A ticket model learns concern classes based on labeled examples with cross-entropy loss.
- Limits
- Double-apply Softmax before PyTorch CrossEntropyLoss; false label types or shapes; ignore class imbalance and data errors.
In context
Questions about this term
Is smaller loss always higher accuracy?
No.
Are logits probabilities?
No.
Does each classification need the same loss variant?
No.
All details and background
For a unique target class, cross entropy penalises a low predicted probability for that class. For multiclass classification, PyTorch CrossEntropyLoss expects unnormalized logits and processes the necessary normalization internally. Training loss and practical classification quality are different sizes.
ELI5 – Explained simply
A self-confident wrong answer gets a stronger correction than a suitable prediction.
Why it matters in practice
Relevant for training classifiers and predicting tokens.
Example
Example scenario: A ticket model learns concern classes based on labeled examples with cross-entropy loss.
Pitfalls and limits
- Double-apply Softmax before PyTorch CrossEntropyLoss; false label types or shapes; ignore class imbalance and data errors.
Origin & history
Based on concepts of information theory and became a standard loss function in training probabilistic classifiers.
Use in mid-sized companies
- Ensure label definition and data quality; additionally evaluate confusion matrix, recall and technical error costs on independent data.
Related terms
- Softmax
- Logits
- Negative log likelihood
- Classification
- Label smoothing
- Accuracy
Matching Flow-alerta service
TrainingAI consulting – harmonize training goals and technically relevant model metrics.Matching tool / use case
View use casesTool: PyTorch CrossEntropyLoss; Use Case: Training of a ticket classifier.Sources
- docs.pytorch.orgPyTorch, CrossEntropyLoss
Quellenprüfung: 09.10.2026. Definition und dokumentierte Einordnung abgeglichen; Beispiele beschreiben mögliche Anwendungen.
Editorially reviewed on: