Training data

What role does training data play?

Training data consists of the examples from which a model learns relationships or parameters during training. The information included depends on the task and learning method. In supervised learning, examples contain inputs and corresponding targets. Reliable assessment also needs data with other roles, such as model selection and final evaluation. The amount of training data alone says little about whether it adequately covers the later task.

In a fictional product project, spare-part photos with reviewed class labels are to be used for training. The team plans images covering relevant views, lighting conditions and backgrounds for the intended use. Many copies of the same photo increase the file count but do not create corresponding variety. Before training, the source, class definitions and division between learning and assessment are documented. This includes recording which images show the same physical specimen or belong to the same capture series. Later assessment can then account for these dependencies instead of unknowingly treating closely related images as independent examples.

Sources

• Google · Datasets: Dividing the original dataset