strategies / sparsity
strategy 03 · sparsity
Populace deliberately assembles a candidate frame larger than any deployable file needs, so it can cover state and congressional-district targets at once. That candidate frame then has to be reduced to a record count that ships and simulates — and which records survive the cut changes how well the final file matches its targets.
L0 regularization with Hard Concrete gates (Louizos et al. 2018) attaches a gate to every candidate record and lets the optimizer drive it toward zero or one, so the retained-record count becomes an explicit optimization variable rather than a fixed draw. Scored on the full Populace US target surface, target-informed selection beats matched random-sampling baselines at the same record budget:
sparsity-paper abstract and Table 5 (Populace commit 558e46c1; three-year pooled ASEC support file, period 2024). The dense no-L0 run was still declining at its 1,500-epoch budget, so the comparison is a finite-optimization result at a matched compute budget, not a claim that dense calibration could not converge lower given more epochs.
Each candidate record carries a stochastic gate that the optimizer can drive to an exact zero or one while staying differentiable in between, making the expected retained-record count a term the optimizer minimizes alongside calibration error, not a separate thresholding step applied afterward.
The raw gated weights score a 9.86% loss on their own — not good enough to publish. Removing the gates and refitting ordinary calibration weights on exactly the selected 57,240 records lowers that to 4.74%: the gates are valuable as a support selector, and the publication weights should be refit after selection.
A soft L2 penalty and a hard per-record weight-ratio cap bound how concentrated the fitted weights can become, so matching the target surface with fewer records cannot be bought by placing arbitrarily large weight on a handful of survivors.
compared against dense no-L0 calibration, random sampling + reweight, and a dense-random scaled sample, all on the same 32,633-target surface and calibration loss
The reported comparison is a single fixed-penalty, single-seed probe on Populace's current full target surface — a proof of concept that informed selection earns its complexity, not a release-ready budget frontier. Sweeping the normalized penalty across multiple retained counts and concentration settings, and a family-level (not row-level) holdout, are named as next work in the paper rather than run here. Target attainment for the file Populace actually ships is tracked separately on the evaluation strategy page.