Google

Recall is the percentage of all actual positive examples in the dev or test set that the classifier correctly labels as positive.

Recall for a Positive-Class Detector

What does the recall of a cat detector measure?

True or False: Recall is the fraction of positive predictions that turn out to be correct.

Recall in a binary classification task

Match each term to its meaning in a wildfire-detection recall example.

Order the steps to compute recall for a fox classifier on the test set.

Question:
Explain why recall alone does not give a complete picture of how well a bird detector performs on a validation set.

Sample answer:
Recall tells us what fraction of the actual bird images were correctly identified as birds. It does not say anything about how often non-bird images were incorrectly labeled as birds. A detector could score very high on recall by marking nearly everything as a bird, which would capture most true bird images but also create many false alarms. Because recall focuses only on the positive class, it cannot by itself show whether the detector is reliable or accurate overall.

Key points:
- Recall = percentage of actual bird images correctly labeled as bird
- Recall does not measure false positives on non-bird images
- A model can raise recall by predicting the bird label too often
- Recall alone is not enough to judge overall performance

Rubric:
Full credit: correctly defines recall, explains that it ignores false positives/non-bird images, and explains how over-predicting birds can inflate recall. Partial credit: gives the definition but does not clearly explain the limitation. No credit: gives an incorrect definition or no explanation of why recall is incomplete.

Why Recall Does Not Fully Describe a Bird Detector

Scenario:
A wildlife lab checks an image model that is supposed to flag photos containing foxes. In the validation set, 150 images really do show a fox. The model marks 90 of those as foxes and fails to identify the other 60.

Prompt:
What recall does this model achieve for fox images, and what does that number mean in practical terms?

Sample answer:
Recall is the number of true fox images that are correctly identified divided by the total number of true fox images. Here, that is 90 out of 150, which equals 0.60, or 60%.

This means the model detects 60% of the fox photos and misses 40% of the ones that actually contain a fox. If the team wants to find as many fox images as possible, this recall is not very strong.

Key points:
- Recall = correctly identified fox images / all actual fox images
- 90/150 = 60% recall
- 40% of the real fox images were missed
- Low recall means many true positives are not being caught

Estimate recall for a fox-image detector

Question:
In one to two sentences, define recall for a bird classifier evaluated on a development or test set.

Sample answer:
Recall is the fraction of all truly bird images in the evaluation set that the classifier labels as bird.

Key points:
- Recall is a fraction or percentage
- It is computed over all actual bird images in the set
- It measures how many of the true birds were found

Rubric:
Full credit: states that recall is the fraction or percentage of all true bird images correctly labeled as bird. Partial credit: partially correct but incomplete. No credit: incorrect or unrelated definition.

Learn Before

Related