Open Internet by MindsNet

Validating Leave-One-Object-Out CV for Pair-Based Models with Limited Data

The author questions the validity of using leave-one-object-out cross-validation (LOO-CV) for pair-based models, such as Siamese networks, when dealing with very few objects. The concern arises because negative pairs in the test set would include at least one known object, potentially biasing the evaluation.

Computing & Technology, Computer Science, Machine Learning