C3 Cube: A Training-Free, Label-Free Neural Network for Arrhythmia Detection

C3 Cube is a biologically inspired spiking neural network architecture that achieves training-free, label-free arrhythmia detection based on the principle of cortical resonance, reaching 100% accuracy on synthetic ECG datasets. The project is sourced from GitHub (ModelingSolver/C3-Cube), is in preprint status, and was released in June 2026. Its core innovation lies in that its discriminative ability emerges entirely from structural dynamics, without the need for training data, optimization processes, or backpropagation.

Traditional deep learning in medical diagnosis relies on large amounts of labeled data, computing resources, and explicit training, which limits its application in resource-constrained scenarios. C3 Cube raises the question: How do biological nervous systems perform classification without sample-by-sample supervision? Mammalian cortex uses intrinsic dynamics (oscillatory rhythms, lateral inhibition, temporal gating) to distinguish signals from noise—can this be instantiated into a computational architecture without a learning phase?

The basic unit of C3 Cube is the HyperNeuron (an artificial cortical column composed of 8 LIF neurons), which simulates three mechanisms: temporal filter (coherent gating to suppress noise), internal clock (oscillator phase encoding), and activity control (lateral inhibition to prevent runaway). 27 HyperNeurons form a 3×3×3 spatial lattice. Input signals are compressed into 4 features (peak amplitude, linear slope, early/late energy) before injection, and classification is achieved via local dynamics triggering cascaded resonance—no learned classification function is involved.

In an MIT-BIH-style synthetic ECG dataset (5 classes, 750 samples, binary classification task), C3 Cube achieved 100% accuracy with zero training (lr=0.0) — zero false negatives and false positives among 150 test samples — matching the performance of a supervised MLP baseline that required 500 training epochs. Resonance scores show clear class separation: normal heartbeats range from 0.00 to 0.22, arrhythmias from 0.88 to 0.99, with no ambiguous regions at the decision threshold of 0.76.

C3 Cube demonstrates that the principles of cortical resonance (temporal gating, phase coupling, lateral inhibition) can achieve medical-grade arrhythmia detection, with performance comparable to supervised MLP under zero training. It challenges the assumption that 'intelligence must come from learning', showcases the value of structure-first approaches, and opens new possibilities for AI applications in resource-constrained environments (edge devices, privacy-sensitive scenarios).

Current limitations: synthetic dataset, limited sample size, manual feature selection, high inference latency. Future directions: validation on real MIT-BIH data, expansion to multi-class tasks, latency optimization (C/FPGA), formalization of the relationship between resonance thresholds and lattice geometry.