The time, computational resources, and power required by any given AI specialist agent depends upon the tasks being performed, the environment, the algorithms employed, and the agent architecture and will vary from training to performance. Teaching a common robotic pick and place task via one-shot learning from demonstration requires roughly equal computational power for training and for performance. Traditional backprop artificial neural networks require much more computation during training* and much less during performance. On the other hand for agent specialties where performance requires substantial inferencing then that may actually be more costly than initial training was.
* I've typically been training A.s.a. with handcrafted curricula rather than using the vast amount of data that systems like GPTs require.
No comments:
Post a Comment