Tasks
Deep-learning tasks differ mainly in the prediction they require and the supervision available. For vision, moving from one label to boxes and then pixels increases spatial detail, annotation cost, and evaluation complexity.
Task Taxonomy
| Task | Typical output | Central question | Common metrics |
|---|---|---|---|
| Classification | One class or a set of labels | What does the input contain? | Accuracy, F1, AUROC |
| Detection | Boxes, classes, and scores | What objects are present, and where? | AP, mAP, box IoU |
| Segmentation | Pixel-aligned class or instance masks | Which label belongs to each pixel? | mIoU, Dice, mask AP |
| Anomaly Detection | Anomaly score and optional heatmap | What differs from the learned normal state? | AUROC, AUPRO, operating-point errors |
Classification, detection, and segmentation are usually supervised with explicit target categories or geometry. Anomaly detection is often one-class or unsupervised, using normal data to define a reference instead of enumerating every failure.
The boundaries are not absolute. A model may classify detected regions, segment each detected instance, or classify an anomaly heatmap into a final accept-or-reject decision. Choose the task formulation that matches the decision the system must make.