A framework proving that an algorithm can learn accurate concepts from limited examples with high probability.