How Fine-Tuning Allows for Effective Meta-Learning