pngwn/self-training-denoising-forgetting

Self-Training: Denoising vs. Forgetting — Reproduction Artifacts

Reproduction artifacts for "Why Self-Training Helps and Hurts" (arXiv:2602.14029): the U-shaped risk curve from the denoising–forgetting trade-off, in overparameterized linear regression (synthetic, spiked covariance) and a small CIFAR-10 deep-learning analogue.

Read REPORT.md first — it states which claims reproduced, with what numbers, and lists all assumptions, deviations, hardware and cost.

Contents

Path	What it is
REPORT.md	Reproduction report with verdicts per claim
synthetic_selftrain.py	Synthetic experiment: Algorithm 1, Thm 3.2 recursion, iGCV (eqs. 11–12, ridgeless limit)
synthetic_results.json	All synthetic results: per-trial risk and iGCV trajectories, theory curves, selected stopping times, seeds
cifar_selftrain.py	CIFAR-10 self-training script (exact payload run on HF Jobs)
cifar/metrics.json	Test accuracy + pseudo-label accuracy per iteration, full config
cifar/ckpt_t{t}/config.json	Configuration for each iterate t=0..8
plots/synthetic_u_curve.png	U-shaped risk: theory vs simulation (paper Fig 1b analogue)
plots/synthetic_decomposition.png	Systematic ↑ / stochastic ↓ decomposition (Fig 3a analogue)
plots/synthetic_igcv.png	iGCV vs true risk (Fig 4b analogue)
plots/synthetic_isotropic_control.png	Isotropic control: monotone, no U (Corollary B.1)
plots/cifar_selftrain_curve.png	CIFAR-10 accuracy over self-training iterations

Companion model repositories

• pngwn/cifar-selftrain-teacher-t0 — initial weak teacher (42.4% test acc)
• pngwn/cifar-selftrain-best-iterate — best iterate, t=4 (46.8%)
• pngwn/cifar-selftrain-final-iterate — final degraded iterate, t=8 (44.7%)

All nine CIFAR checkpoints (t=0..8) are preserved in the storage bucket pngwn/selftrain-artifacts under cifar/ckpt_t{t}/.

Headline numbers

Synthetic (spiked s=25, p=1200): U-shaped risk with optimal iteration matching theory exactly — t* = 5, 2, 1 for ρ = 1.5, 2.0, 2.5; minimum risk within 1–2% of the deterministic recursion; iGCV recovers t* within ±1 iteration in 90–100% of trials. Isotropic control is monotone as predicted.

CIFAR-10 (ResNet-18, n=5000, η=0.4, K=8): 42.4% (teacher) → 46.8% (best, t=4) → 44.7% (final, t=8): improvement through denoising, then degradation through signal forgetting.

Reproducing

python synthetic_selftrain.py   # CPU, a few minutes; writes synthetic_results.json
python cifar_selftrain.py       # needs a GPU; ~45 min on a T4; writes checkpoints + metrics.json

Both scripts are fully seeded; every reported point traces to a saved configuration.