Is zero shot infererce considered as in-context learning?

Is zero shot infererce considered as in-context learning?

Listening to the video, one shot and few shot inferences are indeed in-context learning. The reason being, in the prompt there is a input and output relation demonstrated to the LLM, from which it can learn our task and our expectation of the output format. However the same cannot be said for zero shot inference (as there is no input and output relation to learn from)

Is the conclusion then that zero shot infererce is NOT in-context learning correct?

I think this is correct. When we say “zero-shot inference” as compared to one-shot or few-shot, I think we are just noting that there is no example the LLM can refer from, as compared to a single example or multiple examples. So there is no in-context learning for the model.

In lab 1, we used in-context learning when we saw that the relatively small LLM we used performed poorly on zero-shot inference (we just tried to see if the model could do our task well !), so we tried in-context learning through one-shot and few-shot inference to try and improve the model’s output by giving it example(s).