QiVC-Net: Quantum-Inspired Variational Convolutional Network, with Application to Biosignal Classification

---

Amin Golnari, Jamileh Yousefi, Reza Moheimani, Saeid Sanei. "QiVC-Net: Quantum-Inspired Variational Convolutional Network, with Application to Biosignal Classification"

Amin Golnari^a, Jamileh Yousefi^b, Reza Moheimani^a, Saeid Sanei^c,d
a Faculty of Computer Science, Chemnitz University of Technology, Chemnitz, Germany
^b Shannon School of Business, Cape Breton University, Sydney, NS, Canada
^c Department of Electrical and Electronic Engineering, Imperial College London, London, UK
^d College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam

Link to the preprint version: arXiv:2511.05730

Run this implementation on Google Colab:

Compatibility:

• TensorFlow: 2.19.0
• Keras: 3.10.0
• Python: ≥ 3.12

---

Framework Overview

QiVC-Net introduces the Quantum-inspired Variational Convolution (QiVC) framework, a novel learning paradigm that integrates probabilistic inference, variational optimization, and quantum-inspired transformations into convolutional neural networks.

The core innovation is the Quantum-inspired Rotated Ensemble (QiRE) mechanism, which applies differentiable, low-dimensional subspace rotations to convolutional weights, inspired by unitary evolution in quantum systems. This enables structured, geometry-preserving uncertainty modeling. The framework is instantiated for phonocardiogram (PCG) signal classification, a challenging task marked by noise, inter-subject variability, and class imbalance.

QiVC-Net also features a Reversal Fusion Residual (RFR) block that captures bidirectional temporal dynamics by processing both forward and time-reversed inputs, enhancing robustness and temporal coherence.

---

Key Features:

1. QiVConv Layer:

   • A probabilistic convolutional layer that performs norm-preserving subspace rotations of kernel weights.
   • Enables expressive uncertainty-aware representations.

2. Quantum-Inspired Rotated Ensemble (QiRE):

   • Injects structured stochasticity via unitary-inspired rotations in a learnable low-dimensional subspace.
   • Preserves weight-space geometry and avoids the instability of isotropic Gaussian noise.

3. Uncertainty-Aware Dual-Path Architecture (RFR Block):

   • Processes input signals in both forward and reversed temporal directions.
   • Fuses features via LSTM layers and residual refinement for robust temporal modeling.

4. Robust PCG Classification:

   • Evaluated on PhysioNet/CinC 2016 and CirCor DigiScope 2022 datasets.
   • Achieves 97.84% and 97.89% average accuracy, respectively, with strong calibration (low ECE) and noise robustness.

5. Clinically Aligned Design:

   • Lightweight, interpretable, and efficient, suitable for deployment in safety-critical applications.
   • Uses a composite loss (CCE + Dice) with dynamic weighting to handle class imbalance.

---

Usage of QiVConv in Practice

The QiVConv layer can be used as a drop-in replacement for standard convolutional layers in any TensorFlow/Keras model. It introduces structured uncertainty into the convolutional weights by maintaining separate learnable mean (mu) and standard deviation (rho) parameters for each weight. It supports end-to-end training with standard optimizers. It is particularly well-suited for challenging tasks where uncertainty quantification, robustness to noise, and temporal symmetry are critical.

---

If our work is helpful to you, please kindly cite our paper as:

@article{golnari2025qivc,
  title={QiVC-Net: Quantum-Inspired Variational Convolutional Network, with Application to Biosignal Classification},
  author={Golnari, Amin and Yousefi, Jamileh and Moheimani, Reza and Sanei, Saeid},
  journal={arXiv preprint arXiv:2511.05730},
  year={2025}
}