But what if we drop the word comprehension, and we just go with “functional approximation X -> Y, computed from a finite dataset, which minimizes a predictive r

But what if we drop the word comprehension, and we just go with “functional approximation X -> Y, computed from a finite dataset, which minimizes a predictive risk “? It’s unclear why compression is necessary there, except as a practical benefit.