sample

Pluck a guitar string and the wave that races towards the bridge reflects and comes back. The string then carries two identical waves moving in opposite directions, and what you see is neither of them: a pattern that stays in place and only breathes up and down.

Adding the two waves #

Take two waves of the same amplitude $A$, wavenumber $k$ and angular frequency $\omega$, one moving right and one moving left:

Their sum, by the identity $\sin a + \sin b = 2 \sin\tfrac{a+b}{2} \cos\tfrac{a-b}{2}$, is

The position and the time have separated. Every point of the string moves in step, with an amplitude $2A\sin(kx)$ that depends only on where it is.

Nodes and antinodes #

Where $\sin(kx) = 0$, at $x = n\pi/k$ or every half wavelength, the string never moves: these are the nodes, marked in the figure. Halfway between them the amplitude reaches $2A$, the antinodes. A string of length $L$ fixed at both ends can only hold waves that put a node at each end, so its allowed wavelengths are

which is why a string sounds a fundamental note and its harmonics, and nothing in between.

Gradient descent is the workhorse of machine learning, and it has a well-known weakness. On a long, narrow valley it bounces from wall to wall while creeping along the floor. Adding momentum, a running memory of past gradients, is a one-line change that often makes it several times faster. This article shows why, on the simplest valley there is.

A quadratic valley #

Take a bowl stretched along one axis,

Its gradient is $\nabla f(w) = (\lambda_1 w_1,\ \lambda_2 w_2)$, and the ratio $\kappa = \lambda_2 / \lambda_1 = 10$, the condition number, measures how narrow the valley is. Plain gradient descent takes steps of size $\alpha$:

A step large enough to make progress along the shallow direction overshoots in the steep one, and a step small enough for the steep one crawls along the shallow one. The best fixed step, $\alpha = 2/(\lambda_1 + \lambda_2)$, shrinks the error by a factor of $(\kappa - 1)/(\kappa + 1) \approx 0.82$ per step.

Adding momentum #

Momentum keeps a velocity $z$ that accumulates gradients and decays by a factor $\beta$ each step:

Oscillations across the valley point in alternating directions and cancel in $z$; the small, consistent pull along the floor adds up. With the best choice of $\alpha$ and $\beta$ the error shrinks by $(\sqrt\kappa - 1)/(\sqrt\kappa + 1) \approx 0.52$ per step: about 11 steps instead of 35 to reduce it a thousandfold.

What to try #

• Set $\beta = 0$ and raise $\alpha$ until the path zig-zags, then diverges past $\alpha = 0.2$.
• Keep that $\alpha$ and raise $\beta$: the zig-zag damps, and the path reaches the centre sooner.
• Push $\beta$ towards $1$: too much memory and the path overshoots and circles.

Further reading #

This sample follows the style of Distill articles and borrows its example from Gabriel Goh’s Why Momentum Really Works, which analyses the same valley in far more depth.