Resonant Learner – A Smarter Early Stop for Deep Models (github.com)

Resonant Convergence Analysis (RCA) is a production-ready, open-source early-stopping system that detects true convergence by analyzing oscillation patterns in validation loss rather than relying on simple patience heuristics. Using resonance metrics—β (resonance amplitude) and ω (resonance frequency)—RCA distinguishes stable plateaus from transient stagnation, automatically reduces the learning rate, checkpoints the best model, and halts training when convergence is detected. The authors report 25–47% compute savings (36% on average) across four datasets (BERT SST2, MNIST, CIFAR‑10, Fashion‑MNIST) on an NVIDIA L40S, often preserving or slightly improving final accuracy (e.g., BERT stopped at epoch 7 saving 30% with 92.55% accuracy; CIFAR‑10 saw +1.35%). Technically, RCA is based on log‑periodic resonance analysis: convergence correlates with rising β, shrinking loss amplitude, and stabilization of ω toward an empirical regime. The ResonantCallback for PyTorch supports EMA smoothing, configurable patience_steps, min_delta, up to two LR reductions (factor 0.5), min_lr thresholds, and automatic best‑model restoration. v5 fixes a plateau-detection bug by lowering the β threshold to 0.70. RCA integrates with TensorBoard/W&B, supports multi‑GPU and distributed runs, and is best suited for long, expensive training jobs and hyperparameter searches (less useful for very short runs or cases needing fixed epoch schedules).