← AI Terminology
Early Exit
Early exit attaches intermediate predictors so easy inputs can leave a deep network at shallower layers, saving compute at inference.
Used in vision and NLP cascaded models.
Used in vision and NLP cascaded models.
Why It Matters in AI
Deep models over-spend on easy examples. Early exits cut average latency for classification and some sequence tasks when confidence is high mid-network. Related to adaptive computation ideas in LLMs.
Key Points
| Aspect | Description |
|---|---|
| Use | Mobile vision, efficient NLP classifiers |
| Risk | Wrong early decisions; threshold tuning |
| Benefit | Lower mean latency/energy |
| Related | Mixture of Depths, cascaded models |
| Training | Multi-exit losses jointly |
| Mechanism | Branch classifiers + confidence thresholds |
Simple Analogy
Leaving the exam when you’re sure of answers on page one instead of filling every optional page — save time when confidence is high.
Common Usage Examples
- BranchyNet-style multi-exit CNNs
- Confidence-based exit thresholds
- Measure accuracy vs mean FLOPs
- Avoid exits on safety-critical uncertain cases
Summary
In short: Early exit lets easy examples leave deep networks early — cutting average compute when confidence is high.