Innovative Application of Quantum Neural Networks in MNIST Handwritten Digit Recognition

This project was released by Paradoxical7 on GitHub on May 31, 2026 (link: https://github.com/Paradoxical7/QNN). Its core is building a quantum neural network based on the Qiskit framework, achieving a 96% test accuracy on MNIST handwritten digit recognition with only 24 trainable parameters through an innovative black-and-white amplitude encoding method, exploring the parameter efficiency potential of quantum machine learning.

Quantum Machine Learning (QML) is an emerging field that integrates quantum computing and deep learning. Traditional neural networks require millions of parameters to achieve high accuracy, while Quantum Neural Networks (QNNs) are expected to achieve comparable performance with fewer parameters. As a classic dataset, MNIST contains 60,000 training images and 10,000 test images (28×28 grayscale handwritten digits), and this project is a practice of QNN on this task.

Quantum Neural Network Structure

The QNN consists of three parts: data encoding layer (classical to quantum state conversion), parameterized quantum circuit (trainable quantum operations), and measurement layer (quantum to classical output conversion).

Innovation in Black-and-White Amplitude Encoding

Traditional methods require 784 qubits to process MNIST images. This project uses amplitude encoding: the square of the quantum state amplitude corresponds to the pixel probability distribution, allowing the image to be represented with far fewer qubits than the number of pixels, leveraging quantum parallelism.

Parameter Efficiency

Only 24 trainable parameters, which is a significant reduction compared to classical fully connected networks (tens of thousands of parameters) and lightweight CNNs (thousands of parameters).

Technology Selection

The Qiskit framework is adopted (mature ecosystem, flexible circuit design, compatibility with simulation and real hardware); training uses the parameter-shift rule to calculate gradients, combined with classical optimizers (e.g., Adam) and mini-batch processing.

The project achieved a 96% MNIST test accuracy on a quantum simulator. The key achievement is parameter efficiency: only 24 trainable parameters, which is much lower than traditional models. This proves the feasibility of quantum machine learning algorithms and provides an example for parameter-efficient learning.

Technical Challenges

• Qubit count limitation: unable to process high-resolution images
• Training stability: quantum circuit optimization is prone to local optima
• Simulation cost: classical simulation cost of large-scale quantum systems grows exponentially
• Hardware limitations: noise and error rates of real quantum devices affect performance

Comparison with Classical Models

Classical CNNs can easily achieve 99%+ accuracy on MNIST. The quantum advantage currently lies in parameter efficiency rather than absolute performance; practical deployment needs to consider the availability and cost of quantum hardware.

Conclusion

This project demonstrates the great potential of quantum machine learning—achieving considerable performance with very few parameters, suggesting that future AI may shift to a more intelligent and efficient computing paradigm.

Recommendations

For developers getting started with quantum ML, this project is a good starting point: concise code, using the mature Qiskit framework, and easy to verify results based on the MNIST task. The integration of quantum computing and AI is still in its early stages, but open-source projects are pushing the boundaries and showing the possibilities of future computing.