Neural networks can be trained on which types of datasets?

• A. Labeled datasets only
• B. Unlabeled datasets only
• C. Both labeled and unlabeled datasets
• D. None

Answer: (C)

Neural networks can be trained using data with targets or without targets. Labeled datasets provide input-output pairs, enabling supervised learning where the model learns to map inputs to the correct outputs and tightening the map with a loss that measures error. Unlabeled datasets offer only inputs, so the network learns by uncovering structure—through reconstruction, predicting transformations, or contrasting similar and dissimilar examples—via unsupervised or self-supervised objectives. In many scenarios, you blend both kinds: semi-supervised or self-supervised learning uses large amounts of unlabeled data to learn good representations, then leverages a smaller labeled set to supervise the final task. So, neural networks can be trained on both labeled and unlabeled datasets, depending on the method and goal.