Deep Learning

Accuracy, precision and recall

In deep learning object class predictions are often characterised by the following accuracy metrics.

Accuracy - ratio of correct classifications i.e. "how correct predictions are overall"
Precision - ratio of correct identified positive out of predicted positive i.e. "how correct prediction are of positive"
Recall - ratio of correct predicted positive out of actual positive i.e. "how many of actual positive are found"

Measures are calculated as follows.

                                Predicted 

                            Postive   Negative
 
  Actual /      Positive       TP        FN
Ground Truth    Negative       FP        TN

Accuracy = Accurately predicted / All = TP + TN / All
Precision = True positive / Predicted Postive = TP / (TP + FP)
Recall = True positive / Actul Positive = TP / (TP + FN)

References

Accuracy vs. precision vs. recall in machine learning: what's the difference?