Multiclass classification chooses among more than two classes. In single-label multiclass classification, each input belongs to exactly one class. A Reuters newswire classifier, for example, assigns each article to one of 46 mutually exclusive topics. section 4.2
Set up a 46-topic classifier
For a single-label classifier, use a final Dense layer with 46 units and softmax activation. Match the loss to the targets: categorical_crossentropy for one-hot targets, or sparse_categorical_crossentropy for integer-encoded targets. The number of output units follows the number of topics. section 4.2 section 20.6
Check the labeling task before choosing that setup. If an input can have several labels, it is a multilabel problem instead: use a sigmoid output with binary_crossentropy and K-hot targets, rather than softmax’s single-class constraint. section 4.2 section 20.6
Problem framing comes first: identify the inputs, the available annotations, and whether the desired output is one class or several. Define success metrics and validation splits before settling on an architecture. section 6.1.1 section 20.6
