AI and deep learning researchers, instead of trying to eke out another percentage point on latest perceptual benchmark, consider what architecture you might need to capture this, without extensive training data: New Caledonian crows create compound toolshttps://www.youtube.com/watch?v=xcN3WzeKDlE …
-
How much "training data" has a New Caledonian crow received over the course of its life, constantly interacting with objects and gaining a sense of physical dynamics? Less than an "extensive" amount?
The issue of course how to transfer from a lot of stuff that isn’t obviously task relevant (and also what prior knowledge might be leveraged).
