Deep Learning with PyTorch Quick Start Guide

Evaluating classifiers

You may think that the best way to measure the performance of a classifier is to count the proportion of successful predictions compared with the total predictions made. However, consider a classification task on a dataset of handwritten digits, where the target is all the digits that are not 7. Just guessing that every sample is not 7 will give a success rate, assuming the data is evenly distributed, of 90%. When evaluating classifiers, we must consider four variables:

• TP true positive: The predictions that correctly identify a target
• TN true negative: The predictions that correctly identify a non-target
• FP false positive: Predictions that incorrectly identify a target
• FN false negative: Predictions that incorrectly ...