Given paired data \(\mathcal{D}=\{(x_i,y_i)\}\), estimate the unknown map \(f:\mathcal{X}\to\mathcal{Y}\) that relates them.
It memorizes observed points and does not tell you \(f(x)\) for new \(x\) — classic overfitting.
Family = class of candidates (e.g. neural nets, linear models). Algorithm = how you pick one using data (e.g. backprop, least squares).
Flatten / stack the \(P\times Q\) pixel matrix into a vector in \(\mathbb{R}^{PQ}\).
Gender is an abstract human construct weakly tied to raw reflectance; aircraft dynamics follow relatively simple measurable physics.
Late-1980s statistical speech recognition that ignored detailed vocal-tract physics; it worked via repeated observation + stats — the same spirit as modern ML.
Treat \(X,Y\) as random variables and estimate the relevant distributions (paralleling estimating \(f\)).