Can we change an AI model's architecture without retraining it from scratch?

I may be confused.

How is priming different from transfer learning?