Data-efficiency vs sample-efficiency I noticed that in the context of RL, people call the ability to learn from little data "sample-efficiency". However, in the context of supervised learning, it is called "data-efficiency". Why is that? Is there a subtle difference I'm missing?
 A: Yes they are the same exact thing, some prefer one term over the other.
The idea in supervised learning is, for example, do I really need 6000 samples per class to train a model on MNIST? In a lot of real world problems we don't have the means of labelling that amount of images to create a dataset. Can we somehow get the same results with fewer images? This is the concept of sample efficiency.
In RL you have the same exact motivation, because it is much more "expensive" to do actual simulations in your environment, that to replay your past experiences and try to learn from them as much as you can (e.g. playing a game of pacman might take minutes; it is waaaaay to inefficient to train a model purely on actual pacman games).
A: Although the term "sample-efficient" appears more in reinforcement learning related materials, this seems to be a jargon related issue, and the terms can be used interchangeably.
The following quote is from an ICML workshop, where both terms are used under general machine learning, not particularly RL:

The ability to learn in a sample-efficient manner is a necessity in
these data-limited domains. Collectively, these problems highlight the
increasing need for data-efficient machine learning...

