Robust Quantum Machine Learning Achieves Increased Accuracy on MNIST and FMNIST Datasets

Key Takeaways

• MPS encoding yields low‑depth quantum circuits.
• Reduces qubit count versus one‑qubit‑per‑feature methods.
• Enhances robustness against classical adversarial attacks.
• Demonstrated on MNIST, FMNIST with higher accuracy.
• Validated on superconducting quantum hardware.

Summary

Researchers at the University of Melbourne introduced a Matrix Product State (MPS) based encoding scheme that dramatically reduces quantum circuit depth while preserving classification performance. By iteratively applying singular value decomposition, the method creates low‑depth, approximate encodings that require fewer qubits than traditional one‑qubit‑per‑feature approaches. Experiments on the MNIST and FMNIST datasets showed improved accuracy and heightened resistance to classical adversarial attacks. A small‑scale test on a superconducting quantum processor confirmed the practicality of the technique for near‑term hardware.

Pulse Analysis

Encoding classical data efficiently remains a bottleneck for quantum machine learning (QML). The new method leverages the Matrix Product State formalism, representing high‑dimensional vectors as a chain of low‑rank tensors. By reshaping input vectors and applying singular value decomposition, the authors generate a sequence of small matrices that translate directly into shallow quantum gates. This tensor‑network strategy cuts circuit depth from exponential to linear scaling and trims qubit requirements, addressing both coherence limits and connectivity constraints of current quantum processors.

The practical impact is evident in the reported experiments. Variational quantum classifiers built with MPS‑based encodings achieved state‑of‑the‑art accuracy on the MNIST and FMNIST image benchmarks while showing measurable resilience to classical adversarial perturbations. Crucially, the authors demonstrated the approach on a superconducting device, confirming that the low‑depth circuits can be executed within the coherence windows of today’s hardware. The nearest‑neighbour gate layout and limited CNOT count further align the technique with the architectural realities of NISQ platforms, making it a viable candidate for near‑term deployment in secure data‑analysis pipelines.

Beyond immediate performance gains, the work signals a shift toward noise‑aware algorithm design in the quantum industry. By embracing approximate encodings that exploit inherent quantum noise tolerance, developers can build QML models that are both scalable and robust. Future research will need to explore defenses against quantum‑capable adversaries and extend the methodology to larger, more complex datasets. Nonetheless, the MPS encoding framework provides a concrete, hardware‑friendly pathway for enterprises seeking to integrate quantum‑enhanced analytics into their AI stacks.