A training job keeps running after validation loss stops improving. What saves time?
Why D is the answer
Early stopping
Early stopping ends a job once validation loss stops improving, which saves training time.
In detail
- Early stopping ends the job when validation stops improving.
- It can also limit overfitting.
Use case
An illustrative scenario, not a real incident.
Validation loss stops improving at epoch 12 of 100. Early stopping ends the job there, and saves the remaining time.
See it in different ways
flowchart TD
A{"Results"} -->|"Poor on training"| B["Underfitting: more features or a bigger model"]
A -->|"Good on training, poor on validation"| C["Overfitting: regularize or add data"]stateDiagram-v2 [*] --> Train Train --> Validate: tune Validate --> Train: adjust Validate --> Test: choose the final model Test --> [*]
flowchart TD
Q{"Is validation loss still improving?"}
Q -->|"Yes"| A0["Keep training"]
Q -->|"No"| A1["Stop early"]Key point: Stop when validation stops improving.
Why the other options are wrong
- A. A higher learning rate
- A higher rate does not end the job.
- B. Managed Spot Training
- Spot lowers the price, and does not stop training.
- C. A larger batch size
- Batch size does not end the job.