A vision team finds that early epochs on a mixed-difficulty dataset are unstable. They reorder training so clean, obvious images appear first and ambiguous, noisy images are phased in later, and convergence improves. Which training strategy is this?
Show answer and explanation
Answer: A
Curriculum learning borrows from how people study: present easy or high-confidence examples first so the model settles into a sensible region of parameter space, then introduce harder cases that refine the decision boundary. The schedule can be a fixed ordering by a difficulty score or a self-paced rule that admits harder samples only once the loss has dropped. It often improves stability and final accuracy on noisy or long-tailed data, at the cost of an extra design decision about how difficulty is measured and when to advance.
Why the other options are wrong
- B. Batch normalisation rescales activations per mini-batch; it changes optimisation dynamics but says nothing about the order in which examples are presented.
- C. Data parallelism splits a batch across devices to speed up the same training order; it does not sequence examples by difficulty.
- D. Early stopping halts training when validation loss stops improving, truncating a run rather than ordering the data inside it.