sqdr-cnn-spiking-quantum

name	sqdr-cnn-spiking-quantum
description	SQDR-CNN methodology — joint training of convolutional SNNs and quantum circuits with surrogate gradient and quantum data-reupload for parameter-efficient hybrid models.
category	ai_collection

SQDR-CNN: Spiking-Quantum Data Re-upload Convolutional Neural Network

Overview

SQDR-CNN is a parameter-efficient hybrid architecture that enables joint training of convolutional Spiking Neural Networks (SNNs) and quantum circuits within a single backpropagation framework, using surrogate gradients and quantum data-reupload.

Source: arXiv:2512.03895 (Published in PeerJ Computer Science, 2026) Authors: Luu Trong Nhan, Luu Trung Duong, Pham Ngoc Nam, Truong Cong Thang

Core Methodology

1. Joint Backpropagation Framework

• Unlike prior SQNN implementations requiring pretrained SNN encoders, SQDR-CNN trains end-to-end
• Convolutional SNN encoder + quantum circuit share gradients
• Surrogate gradient technique makes non-differentiable spiking activity trainable
  • Uses smooth approximation (e.g., sigmoid, fast sigmoid) of Heaviside spike function
  • Gradient flows through spike generation during backprop

2. Quantum Data Re-upload

• Input data is encoded into quantum circuit multiple times across layers
• Each re-upload layer applies:
  1. Data encoding (rotation gates based on input features)
  2. Variational ansatz (parameterized entangling gates)
  3. Measurement (expectation values as output)
• Theoretical advantage: single-qubit circuit with N re-uploads ≈ N-qubit expressivity

3. Architecture Design

Input → ConvSNN (spike encoder) → Flatten spikes → 
  Data Re-upload Layer 1 (encode + variational) → 
  Data Re-upload Layer 2 (encode + variational) → 
  ... → 
  Measurement → Classification

4. Noise-Robust Training

• Evaluate under noisy simulated quantum environments
• Test different training algorithm-initialization combinations
• Deploy on actual quantum hardware (IBM Q)

Performance Results

• 86% of SOTA SNN baseline accuracy with only 0.5% of parameters
• Converges without pretrained spiking encoders
• Works without dataset subsetting
• Robust under noisy quantum simulation

Implementation Steps

1. ConvSNN encoder design:

   • Convolutional layers with spiking neurons (LIF or Izhikevich)
   • Surrogate gradient function for backprop (e.g., straight-through estimator)
   • Temporal dimension: simulate over T timesteps

2. Quantum circuit design:

   • Choose number of qubits based on flattened spike features
   • Design data re-upload schedule (how many times to re-encode)
   • Select variational ansatz (RY-RZ-CNOT ladder, etc.)

3. Hybrid training loop:

   • Forward: SNN → spikes → quantum encoding → measurement
   • Backward: classical loss → quantum gradient (parameter-shift) → SNN surrogate gradient
   • Use hybrid optimizers (Adam for classical, gradient descent for quantum)

4. Noise simulation:

   • Add depolarizing, dephasing, or amplitude damping noise
   • Use IBM Qiskit noise models for realistic simulation

Pitfalls

• Surrogate gradient choice: Different functions (sigmoid, arctan, triangle) significantly affect training stability
• Temporal steps: Too few → poor SNN dynamics; too many → slow training
• Data re-upload depth: More layers = more expressivity but also more noise sensitivity
• Qubit-feature matching: Number of features may exceed available qubits — use dimensionality reduction
• Pretraining dependency: Prior SQNN required pretrained SNN; SQDR-CNN removes this but needs careful initialization
• Hardware deployment: Real quantum devices have limited qubits and high error rates

Activation Keywords

spiking quantum neural network, SQNN, surrogate gradient, quantum data-reupload, hybrid quantum-classical, SNN backpropagation, convolutional SNN, parameter-efficient, neuromorphic quantum, joint training, SQDR-CNN, noise-robust quantum

References

• arXiv:2512.03895 — Parameter efficient hybrid spiking-quantum convolutional neural network with surrogate gradient and quantum data-reupload
• PeerJ Computer Science 12 (2026): e3554