From pixels to class scores
A model receives an image representation and returns scores for possible classes. A softmax can turn scores into a distribution across mutually exclusive classes, but high probability does not guarantee correctness.
Training and evaluation
Labeled examples can teach the model to associate images with classes. Evaluate on images not used to train or tune it. A test set should resemble the intended use as closely as possible. Change the cat score while the dog score remains fixed. The two displayed probabilities compete for the same total of one.
INTERACTIVE EXPLORATION
Compare class scores
A toy classifier gives cat a score you control and dog a fixed score of 0. Softmax turns the two scores into probabilities.
What is missing?
A single image label does not identify where an object is. Object detection adds locations; segmentation assigns labels to pixels or regions. Our interactive example only changes scores, not the pixels or a trained model.
Check your understanding
Does a label “cat” tell us where the cat is in an image?
No. A classification label alone does not provide an object location.