Interactive demo · Foundations of Deep Learning · Chapter 2

Change of variables: the Jacobian and the moving mode

If x = g(y), probability in a thin interval is conserved, so py(y) = px(g(y)) |g′(y)|. The extra factor is what separates a density from an ordinary function: plug g(y) into px without it and the curve no longer matches a histogram of transformed samples, and the maximum of the true density sits somewhere else. In several dimensions the factor becomes |det J|, the local ratio of volumes, which is the quantity normalizing flows (Chapter 18) are built around.

View

Mean μ of px

Standard deviation σ

Map y = g−1(x)

Steepness a of the sigmoid

Shift c

Samples

tanh gain a in (2.78)–(2.79)

Shear b in b x13

Spread s of px = N(0, s2I)

Click or tap the left panel to move the yellow probe square.

Scoreboard

Try these

In the lecture notes