Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/jeragilo/noise-robust-hqnn-framework

Master’s thesis research framework for noise-robust hybrid quantum–classical neural networks (HQNNs), evaluating reliability, architecture design, and deployment strategies on NISQ hardware.
https://github.com/jeragilo/noise-robust-hqnn-framework

cirq hqnn hybrid-ai hybrid-models hybrid-quantum-classical-algorithms masters-thesis nisq noise-robust-learning pennylane qiskit quantum-computing quantum-machine-learning quantum-neural-networks research research-framework variational-algorithms

Last synced: 4 days ago
JSON representation

Master’s thesis research framework for noise-robust hybrid quantum–classical neural networks (HQNNs), evaluating reliability, architecture design, and deployment strategies on NISQ hardware.

Awesome Lists containing this project

README 

          
# Noise-Robust Hybrid Quantum Neural Networks Framework



This repository contains the experimental codebase for my Master’s thesis:



**Noise-Robust Hybrid Quantum Neural Networks: A Framework for Scalable Quantum AI in the NISQ Era**



The project investigates the reliability, robustness, and practical limitations of hybrid quantum–classical neural networks under simulated NISQ-era noise.



This repository is structured as a reusable benchmarking and optimization framework, not only as a collection of independent demos.



---



## Flagship Result: Stability-Regularized Multi-Observable HQNN



The strongest configuration combines:



- Stability-regularized noise-aware training

- Multi-observable quantum feature extraction

- Linear HQNN architecture

- Random Forest learned classical readout



Result under depolarizing noise:



| Metric | Value |

|---|---:|

| Clean accuracy | 0.9600 |

| Noisy accuracy | 0.9600 |

| Accuracy drop | 0.0000 |

| Robustness score | 1.0000 |

| Gain over random parity baseline | +0.5067 |



This supports the main thesis claim that HQNN robustness improves when noise awareness is embedded into the optimization process and richer quantum measurement information is preserved through learned hybrid readouts.

### Multi-Channel Noise Validation



The flagship HQNN configuration was also validated across four NISQ-relevant noise channels at eval_noise = 0.05.



| Noise Channel | RF Clean Accuracy | RF Noisy Accuracy | Accuracy Drop | Robustness Score |

|---|---:|---:|---:|---:|

| Depolarizing | 0.9600 | 0.9467 | 0.0133 | 0.9861 |

| Bit flip | 0.9467 | 0.9467 | 0.0000 | 1.0000 |

| Phase flip | 0.9533 | 0.9533 | 0.0000 | 1.0000 |

| Amplitude damping | 0.9600 | 0.9267 | 0.0333 | 0.9653 |



This strengthens the thesis claim by showing that the flagship HQNN method generalizes beyond a single depolarizing-noise condition.

---



## Main Thesis Claim



Naive HQNN models with fixed parity readout can perform poorly because they compress quantum measurement information into a single decision signal. This framework shows that HQNN performance can be substantially improved by combining:



1. Noise-aware training objectives  

2. Stability regularization  

3. Multi-observable quantum feature extraction  

4. Architecture search  

5. Learned classical readouts  

6. Repeated-trial and statistical validation  



The framework moves the thesis from simple HQNN benchmarking toward a reusable method for discovering noise-robust hybrid quantum-classical learning configurations.



---



## Key Algorithmic Contributions



### 1. Learned Classical Readout HQNN



Instead of using a fixed parity threshold, the circuit measurement distribution is used as a quantum feature representation for a learned classical readout.



Best observed learned-readout result:



| Method | Clean Accuracy | Noisy Accuracy |

|---|---:|---:|

| Fixed parity readout | 0.3444 | 0.3667 |

| Learned logistic readout | 0.8333 | 0.8222 |



---



### 2. Multi-Observable HQNN Readout



The framework extracts richer quantum-derived features:



- Full bitstring probabilities

- Single-qubit Z expectations

- Pairwise ZZ correlations

- Global parity expectation

- Probability-distribution statistics



This produced a 31-dimensional quantum-derived feature vector.



Best observed multi-observable result:



| Method | Clean Accuracy | Noisy Accuracy |

|---|---:|---:|

| Multi-observable Logistic Regression | 0.8000 | 0.8200 |

| Multi-observable Random Forest | 0.8733 | 0.8533 |



---



### 3. Architecture Search for HQNN Robustness



The framework compares different entanglement structures:



- No entanglement

- Linear entanglement

- Ring entanglement

- Full entanglement



Best architecture-search result:



| Architecture | Readout | Clean Accuracy | Noisy Accuracy |

|---|---|---:|---:|

| Linear | Random Forest | 0.8933 | 0.8867 |



---



### 4. Best-Architecture Noise Sweep



The optimized linear HQNN configuration was tested across increasing depolarizing noise levels.



| Noise Level | Accuracy |

|---:|---:|

| 0.00 | 0.8467 |

| 0.01 | 0.8467 |

| 0.03 | 0.8467 |

| 0.05 | 0.8467 |

| 0.07 | 0.8200 |

| 0.10 | 0.8133 |



This shows that the optimized configuration remains stable across moderate noise levels.



---



### 5. Dual-Loss and Stability-Regularized HQNN Training



The framework implements noise-aware objective functions that combine clean behavior, noisy behavior, and stability under perturbation.



Training objectives include:



- Standard clean-loss training

- Noise-aware training

- Dual-loss training

- Stability-regularized training



Best observed result:



| Training Mode | Readout | Clean Accuracy | Noisy Accuracy | Accuracy Drop |

|---|---|---:|---:|---:|

| Stability-regularized | Multi-observable Random Forest | 0.9600 | 0.9600 | 0.0000 |



---



## Framework Overview



The framework uses:



- Qiskit

- Qiskit Aer

- Qiskit Machine Learning

- Cirq

- PennyLane

- scikit-learn

- NumPy

- Matplotlib



The contribution of this project is the reusable evaluation, optimization, reporting, and robustness-analysis layer built around those libraries.



---



## Framework Features



### Standardized Dataset Handling



- Synthetic classification data

- Iris

- Wisconsin Diagnostic Breast Cancer dataset

- Quantum-compatible low-dimensional preprocessing



### Noise-Analysis Toolbox



Supported noise models:



- Depolarizing noise

- Bit-flip noise

- Phase-flip noise

- Amplitude damping



### Robustness Metrics



The framework includes reusable metrics:



- `accuracy_drop`

- `robustness_score`

- `degradation_slope`

- `training_instability`

- `cross_framework_deviation`



### Benchmark Pipelines



The framework includes pipelines for:



- Hybrid vs classical comparison

- Noise robustness

- Cross-framework validation

- Learned-readout HQNN evaluation

- Multi-observable HQNN evaluation

- Architecture search

- Best-architecture noise sweep

- Repeated-trial validation

- Statistical validation

- Dual-loss noise-aware training

- Stability-regularized multi-observable HQNN



### Standardized Outputs



The framework generates:



- JSON summaries

- CSV summaries

- Accuracy plots

- Noise curves

- Heatmaps

- Statistical validation reports



---



## Repository Structure



```text

framework/

  datasets.py

  noise_channels.py

  robustness_metrics.py

  reporting.py

  benchmark_runner.py



pipelines/

  main_hybrid_vs_classical.py

  main_noise_robustness.py

  main_cross_framework_validation.py

  main_full_benchmark_summary.py

  main_framework_capabilities_report.py

  main_training_mode_comparison.py

  main_learned_readout_hqnn.py

  main_multi_observable_hqnn.py

  main_architecture_search_hqnn.py

  main_best_architecture_noise_sweep.py

  main_best_architecture_repeated_trials.py

  main_statistical_validation.py

  main_dual_loss_noise_aware_hqnn.py

  main_dual_loss_multi_observable_hqnn.py



demos/

  core/

  industry/



results/

  framework/



env/

  requirements.txt



run_framework.py

Running the Framework



Run the full framework:



python run_framework.py



Run individual framework pipelines:



PYTHONPATH=. python pipelines/main_hybrid_vs_classical.py

PYTHONPATH=. python pipelines/main_noise_robustness.py

PYTHONPATH=. python pipelines/main_cross_framework_validation.py

PYTHONPATH=. python pipelines/main_full_benchmark_summary.py

PYTHONPATH=. python pipelines/main_framework_capabilities_report.py



Run advanced HQNN optimization pipelines:



PYTHONPATH=. python pipelines/main_learned_readout_hqnn.py

PYTHONPATH=. python pipelines/main_multi_observable_hqnn.py

PYTHONPATH=. python pipelines/main_architecture_search_hqnn.py

PYTHONPATH=. python pipelines/main_best_architecture_noise_sweep.py

PYTHONPATH=. python pipelines/main_best_architecture_repeated_trials.py

PYTHONPATH=. python pipelines/main_statistical_validation.py

PYTHONPATH=. python pipelines/main_dual_loss_noise_aware_hqnn.py

PYTHONPATH=. python pipelines/main_dual_loss_multi_observable_hqnn.py

Demonstration Ecosystem



The project includes a 13-demo experimental ecosystem across Qiskit, Cirq, and PennyLane.



Core Demos

HQNN Toy Classifier

VQE Energy Minimization

QAOA MaxCut

QSVM Anomaly Detection

Noise-Robust HQNN

Cross-Framework Noise Benchmark

Cross-Platform Parity Consistency

HQNN Training Loop with SPSA

Industry-Inspired Demos

Medical Risk Classification

Energy Grid Optimization

Cybersecurity Anomaly Detection

HQNN Explainability

Cross-Noise Robustness Heatmap



Run an individual demo from the repository root:



PYTHONPATH=. python demos/core/demo05_hqnn_noise_robust_qiskit.py



Cybersecurity demo:



PYTHONPATH=. python demos/industry/demo11_cyber_anomaly_qiskit.py

Environment Setup

conda create -n hqnn python=3.11 -y

conda activate hqnn

pip install -r env/requirements.txt

Thesis-Relevant Interpretation



The results suggest that HQNN performance is not determined only by the quantum circuit itself. It depends strongly on the full hybrid pipeline:



how noise is inserted into training,

how measurement information is extracted,

how the classical readout interprets quantum features,

how entanglement architecture is selected,

and how robustness is validated statistically.



The strongest result shows that a stability-regularized, multi-observable, learned-readout HQNN can preserve performance under simulated depolarizing noise, achieving 0.9600 clean accuracy and 0.9600 noisy accuracy in the reported experiment.



Status



The repository currently includes:



Framework layer

Pipeline layer

Core demos

Industry demos

Benchmark outputs

Advanced HQNN optimization pipelines

Statistical validation outputs

Thesis-ready figures and JSON summaries

Contact



GitHub: https://github.com/jeragilo/