Resonant Learner (Community Edition) introduces a new approach to training stabilization — Resonant Convergence Analysis (RCA) — which replaces traditional “patience + min_delta” early-stopping with a dynamic, frequency-based feedback loop.
Instead of watching loss values alone, RCA measures the β-amplitude and ω-frequency of validation oscillations to detect when learning transitions from “searching” to “settling”.
The result: models converge faster with fewer epochs while maintaining (or improving) accuracy.
Highlights
Up to 2× faster training across CIFAR-10, MNIST, and BERT SST-2
Plug-and-play PyTorch callback (no framework changes)
Runs on Windows, Linux, and RunPod GPU (CUDA 12.4+)
Thanks for checking this out! Here are a few quick details for anyone who wants to reproduce the results:
Setup
git clone https://github.com/Freeky7819/resonant-learner cd resonant-learner pip install -U pip setuptools wheel pip install -e . python verify_installation.py
Run examples
# CIFAR-10 baseline python examples/cifar10_rca.py --baseline
# CIFAR-10 with RCA python examples/cifar10_rca.py --use-rca 1
# BERT SST-2 baseline python examples/hf_bert_glue.py --task sst2 --baseline
# BERT SST-2 with RCA python examples/hf_bert_glue.py --task sst2 --use-rca 1
Typical results (Community Edition)
Dataset Baseline Epochs RCA Epochs Δ Compute Accuracy CIFAR-10 60 33 −45 % +0.6 % BERT SST-2 10 6 −40 % +0.4 % MNIST 20 12 −40 % ≈ same
The RCA module is open-core (MIT). SmartTeach + AutoCoach + Stillness are part of the upcoming Pro edition (meta-learning and damping layers).
Happy to answer questions, benchmark requests, or implementation details.
Resonant Learner (Community Edition) introduces a new approach to training stabilization — Resonant Convergence Analysis (RCA) — which replaces traditional “patience + min_delta” early-stopping with a dynamic, frequency-based feedback loop.
Instead of watching loss values alone, RCA measures the β-amplitude and ω-frequency of validation oscillations to detect when learning transitions from “searching” to “settling”. The result: models converge faster with fewer epochs while maintaining (or improving) accuracy.
Highlights
Up to 2× faster training across CIFAR-10, MNIST, and BERT SST-2
Plug-and-play PyTorch callback (no framework changes)
Runs on Windows, Linux, and RunPod GPU (CUDA 12.4+)
Open-source under MIT license
Repo: https://github.com/Freeky7819/resonant-learner