Dynamic Few-Shot Visual Learning without Forgetting