Quantum Hamiltonian Embedding of Images for Data Reuploading Classifiers

Technical Summary: Quantum Hamiltonian Embedding of Images for Data Reuploading Classifiers

Problem Statement
The application of quantum computing to machine learning (QML) has traditionally focused on achieving quantum advantage through runtime speedups, often by "quantising" classical algorithms (e.g., using quantum linear algebra for subroutines like qPCA). However, recent discourse questions whether speedup is the optimal goal for QML, suggesting instead that unique quantum properties or heuristics from classical deep learning should guide model design. A specific challenge arises in image classification: classical deep learning relies heavily on the locality of information and translation invariance (e.g., CNNs, Vision Transformers). Existing QML approaches for images often struggle with high-dimensional data, frequently requiring classical preprocessing (dimensionality reduction via PCA or pre-trained networks) or using embedding methods (like angle or amplitude embedding) that may introduce unwanted biases or fail to preserve the intrinsic two-dimensional structure of image data.

Methodology
The authors propose a Quantum Neural Network (QNN) classifier that integrates the Data Reuploading Circuit architecture with a Quantum Hamiltonian Embedding method.

1. Hamiltonian Image Embedding: Instead of flattening images or encoding pixel values as rotation angles (angle embedding), the model treats the grayscale image matrix $M$ as a Hermitian operator. The image is first "Hermitianised" ($H_M = (M + M^T)/2$) and then encoded into the quantum state via the unitary time-evolution operator $W(t; M) = e^{-iH_M t/2}$, where $t$ is a trainable parameter. This approach preserves the 2D structure of the image and applies a matrix polynomial nonlinearity to the entire image level, avoiding the bias associated with periodic trigonometric functions in angle embedding.
2. Data Reuploading: The model utilizes a variational circuit where the data encoding unitary ($W$) and a parameterized unitary layer ($V$) are repeated $L$ times. The state is prepared as $|\phi(t, \vec{\omega}; M)\rangle = \prod_{i=1}^L [V(\omega_i)W(t_i; M)] |+\rangle^{\otimes n}$. This mimics the reuse of input data in classical deep networks, allowing the model to learn complex functions through repeated interactions between the data and trainable parameters.
3. Architecture: The parameterized layers $V$ consist of $SU(4)$ unitary gates arranged in a brick-wall layout (or $SU(N)$ for larger qubit counts). Classification is performed by measuring projection operators corresponding to class labels, with training optimized via cross-entropy loss.

Key Contributions

• Model Construction: The paper constructs a novel QML classifier combining Hamiltonian embedding with the data reuploading paradigm.
• Performance Benchmarking: The authors demonstrate through numerical experiments that this model significantly outperforms the baseline Quantum Convolutional Neural Network (QCNN) on image classification tasks.
• Design Principles: Based on the model's design and empirical results, the authors propose six guiding principles for future QML model design, shifting the focus from pure speedup to architectural heuristics and data-intrinsic structures.

Experimental Results
The model was evaluated on four datasets: Kaggle CT Medical Images, a subset of the Sklearn Digits dataset, and subsets of the MNIST and FashionMNIST datasets.

• Performance Gains: The proposed "HamEmb" model consistently outperformed the baseline QCNN (which uses amplitude embedding and flattened inputs).
  • On the MNIST test set (8 classes), the HamEmb model achieved 89.72% accuracy compared to the QCNN's 47.03%, a margin of over 40%.
  • On the FashionMNIST test set, HamEmb achieved 77.77% accuracy versus QCNN's 42.90%.
  • On the Sklearn Digits dataset, HamEmb reached 94.40% accuracy against QCNN's 78.95%.
• Robustness: The HamEmb model showed consistent performance across 20 (for smaller datasets) or 5 (for larger datasets) different random parameter initializations, whereas the QCNN performance was more variable and generally lower.
• Scalability Observations: While the model performed well, the authors noted a performance drop when moving from MNIST to the more complex FashionMNIST dataset. They attribute this to the model's global encoding (treating the whole image as one Hamiltonian), which may lack the fine-grained local feature extraction capabilities of smaller convolutional kernels or patch-based approaches.

Significance and Claims
The paper claims that its primary contribution is not a runtime speedup over classical methods, but rather a demonstration that heuristics and empirical knowledge from classical deep learning can be successfully integrated into QML to improve performance.

The authors argue that:

1. Speedup is not the primary metric: Designing QML models should first prioritize performance metrics (like accuracy) and structural alignment with data, rather than seeking quantum speedups.
2. Data Structure Matters: Preserving the intrinsic structure of data (e.g., 2D spatial structure of images) via Hamiltonian embedding is superior to methods that flatten data or introduce periodic biases.
3. Minimize Classical Preprocessing: The model avoids offloading feature extraction to classical backbones (like ResNet), allowing the quantum circuit to process the raw (or minimally processed) data directly.
4. Nonlinearity Sources: The model leverages the nonlinearity inherent in the time-evolution of a quantum Hamiltonian (via the Taylor expansion of the exponential map) rather than relying solely on measurement-induced nonlinearity or classical activation functions.
5. Avoid Unwanted Bias: By avoiding angle embedding, the model sidesteps the bias toward decision boundaries formed by periodic functions.

Ultimately, the paper posits that these six principles—ranging from "Less Initial Focus on Speedups" to "Be Careful of Unwanted Bias"—provide a framework for designing more effective quantum neural networks that respect both the nature of the data and the constraints of current quantum hardware.