out-of-distribution
term_id: out_of_distribution
Category: basic_concepts
Definition
Out-of-distribution (OOD) detection identifies inputs that fall outside the scope of the training data distribution. Models often perform poorly or confidently incorrectly on OOD data, leading to unreliable predictions in real-world scenarios. Detecting these anomalies is crucial for safety-critical applications like autonomous driving or medical diagnostics, ensuring the system recognizes when it lacks sufficient knowledge to make a safe decision.
Summary
Data points that differ significantly from the distribution seen during the model’s training phase.
Key Concepts
- Distribution shift
- Anomaly detection
- Generalization failure
- Safety
Use Cases
- Autonomous vehicle safety checks
- Fraud detection systems
- Medical image analysis validation