Training

Curriculum Learning

Curriculum learning orders training data from simple to complex. Models learn fundamentals first, then tackle harder problems. This mirrors human learning. Curriculum learning can improve convergence speed and final performance. It requires defining example difficulty. Curriculum learning is particularly useful for reasoning and complex tasks. Opposite of random shuffling. Recent work explores automatic curriculum construction.