2 comments

  • Harmonic_Logos 8 hours ago
    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.

  • Harmonic_Logos 8 hours ago
    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