Cross-validation is a statistical resampling method used to estimate the skill of a machine learning model on unseen data. When a dedicated test set is unavailable, cross-validation partitions the available training data into multiple complementary subsets or "folds." The model is trained on a portion of these folds (the training set) and validated on the remaining fold (the validation or "holdout" set). This process is repeated multiple times, with each fold serving as the validation set once. The performance results are then averaged to provide a more robust estimate of the model's generalization ability than a single train-validate split.

B. Randomized controlled trial: This is an experimental design methodology, primarily from medical and social sciences, used to assess the causal impact of an intervention, not to evaluate a predictive model's performance.

C. Average entropy approximation: Entropy measures uncertainty and is a component of loss functions (e.g., cross-entropy loss) used during model training, but it is not a standalone method for evaluating overall model performance.

D. Greedy decoding: This is an inference algorithm used by sequence-generating models to produce an output. It is a method for using a trained model, not for evaluating its performance.

1. University Courseware:

Ng, A., & Niu, G. (2020). CS229 Machine Learning Course Notes: Model Selection and Cross-Validation. Stanford University. In Section 2, "Cross-validation," the document states, "To get a better estimate of generalization error, we can use a procedure called cross validation... In k-fold cross validation, we partition the training set into k disjoint subsets." This directly supports using cross-validation to estimate performance without a separate test set.

2. Peer-reviewed Academic Publications:

Arlot, S., & Celisse, A. (2010). A survey of cross-validation procedures for model selection. Statistics Surveys, 4, 40-79. https://doi.org/10.1214/09-SS054. The paper's introduction (Section 1, p. 41) describes cross-validation as a method "to estimate the risk of an estimator," which is synonymous with assessing its performance on new data.

3. Official Vendor Documentation:

NVIDIA TAO Toolkit 5.2.0 Documentation. (2024). DetectNetv2: K-Fold Cross-Validation for Object Detection. In the section on "K-fold cross-validation workflow," the documentation describes the process of splitting a dataset into k folds for training and validation, which is the core mechanism for evaluating models when a dedicated test set is not used for every training iteration.