Deep Double Descent: Which Regime Is It In?

The problem, from first principles

Double descent has three regimes. Below the interpolation threshold (parameters fewer than samples), test error is U-shaped. At the threshold it peaks. Beyond it, error falls again, so bigger models can generalize better.

From theory to code

Implement double_descent_regime(n, p), returning the regime name for n samples and p parameters.

Constraints

• The interpolation threshold is exactly p == n.

Hints