Which machine learning approach defers computation until a prediction is requested?

The correct answer is lazy learning, which is a machine learning approach that focuses on storing training data and only performing computations when a prediction is requested. This means that during the training phase, a lazy learning model does not build a general model or function from the data. Instead, it keeps all the training instances available for reference at the time of prediction.

When a new query is presented, the model evaluates the stored data to make a prediction. Typical examples of lazy learning include algorithms like k-Nearest Neighbors (k-NN), where the model looks at the closest training examples to make a decision based on their labels at the time of need. This approach can be particularly beneficial in scenarios where the data is large and computational resources can be reserved for when they are truly necessary.

While batch learning involves processing all the training data at once before making predictions, lazy learning allows for on-demand computation. Reinforcement learning and online learning have different methodologies, focusing instead on continuous learning and adaptation over time, which does not align with the deferred computation aspect of lazy learning.