Quantum AI Generates Images with a Novel Hybrid Computer Architecture

Researchers are tackling the limitations of quantum generative models in complex image creation, specifically addressing challenges in scalability and representing multi-modal distributions. Jeongbin Jo, Santanam Wishal from Asia Cyber University, and Shah Md Khalil Ullah from Khulna University of Engineering and Technology, alongside Shan Kowalski from Eleven Dimension LLC and Dikshant Dulai, present a Hybrid Quantum-Classical U-Net architecture incorporating Adaptive Non-Local Observables (ANO). This innovative approach compresses classical data into a quantum latent space and employs trainable observables to capture non-local features, thereby improving generative performance. Their work demonstrates the successful generation of structurally coherent images from the MNIST dataset, suggesting a viable route towards overcoming mode collapse and advancing quantum-enhanced image generation despite current hardware restrictions. By leveraging trainable observables, the model effectively complements classical computation and improves generative performance.

Central to this breakthrough is the exploration of Skip Connections, which play a crucial role in preserving semantic information throughout the reverse diffusion process. Experimental validation using the full MNIST dataset, encompassing digits zero through nine, demonstrates the architecture’s capacity to generate structurally sound and readily identifiable images across all classes.

While current quantum hardware imposes constraints on achievable resolution, the findings establish a viable pathway for mitigating mode collapse and boosting generative capabilities within the constraints of Noisy Intermediate-Scale Quantum (NISQ) technology. The proposed system employs a quantum diffusion model, building upon classical diffusion processes that progressively add noise to data before learning to reverse this process for image creation.

The quantum component introduces a forward process defined by a depolarizing channel, followed by a quantum denoising reverse process. Training is optimized using a loss function based on fidelity or trace distance, ensuring the generated images align with the desired characteristics of the training data.

This approach allows for efficient sampling and generation of complex images by leveraging the unique properties of quantum computation. Furthermore, the research details various quantum data encoding methods, including basis, amplitude, angle, phase, and dense angle encoding, each offering different trade-offs in terms of resource utilization and expressiveness.

The team also investigated general and specialized variational ansatzes, including the Optimal General Two-Qubit Ansatz and 1D Cluster State Mixing, to optimize the quantum circuit’s performance. The study compressed classical data representing the full MNIST dataset, comprising digits 0-9, into a dense quantum latent space to facilitate efficient processing and feature extraction.

Trainable observables were then implemented to capture non-local features, complementing the capabilities of classical processing techniques and addressing limitations in expressibility. This work leveraged a U-Net structure, a convolutional neural network known for its effectiveness in image segmentation and generation, and combined it with quantum circuits to enhance its generative potential.

Skip connections were incorporated within the U-Net to preserve crucial semantic information during the reverse diffusion process, a technique inspired by denoising diffusion probabilistic models. This adaptability is crucial for extracting relevant features from the complex Hilbert space and improving the quality of generated images.

Performance was evaluated by assessing the structural coherence and recognizability of the generated digit images, demonstrating the architecture’s ability to produce meaningful outputs across all ten digit classes. The work explored several quantum data encoding methods including Basis Encoding, Amplitude Encoding, Angle Encoding, Phase Encoding, and Dense Angle Encoding to optimise data representation within the quantum system.

Investigations into Ansatzes for Quantum Machine Learning and Vision, alongside the Optimal General Two-Qubit Ansatz, facilitated the development of effective quantum circuits. Global information propagation was enhanced through the implementation of 1D Cluster State Mixing, improving the model’s ability to process complex data.

Circuit benchmarking was performed using KL Divergence to assess expressibility and the Meyer-Wallach Measure to quantify entangling capability. These metrics provided insights into the model’s capacity to represent complex functions and create quantum correlations. Visualisation of the state space further aided in understanding the model’s internal workings and performance characteristics.

The architecture’s performance was evaluated through analysis of the generative process dynamics and assessment of multi-class generation scalability. Skip Connections were investigated for their role in preserving semantic information during the reverse diffusion process, contributing to the overall quality of generated images.

Findings suggest that hybrid architectures, incorporating adaptive measurements, offer a viable route to mitigate mode collapse and improve generative capabilities within the current NISQ era. This new approach addresses challenges in scalability and expressibility when generating complex, multi-modal data distributions. The model compresses classical data into a dense quantum latent space and uses trainable observables to identify and utilise non-local features, enhancing the generative process.

Experimental results using the full MNIST dataset of handwritten digits demonstrate the architecture’s ability to generate coherent and recognisable images across all digit classes. Skip connections within the network play a crucial role in preserving important semantic information during the reverse diffusion process, which reconstructs images from noise.

Although current quantum hardware limitations restrict the achievable resolution of generated images, the findings indicate that hybrid architectures with adaptive measurements offer a viable route to overcome mode collapse and improve generative performance. The authors acknowledge that the resolution of generated images is currently limited by the capabilities of available quantum hardware.

Future research will likely focus on overcoming these hardware constraints and exploring the potential for further enhancing the model’s expressibility and scalability. This work establishes a promising pathway for leveraging quantum computation to improve generative modelling, particularly in scenarios where capturing complex, non-local relationships within data is crucial.