Quantum Annealing Trains Generative Models Using 2000 Qubits

A new method enhances the generative capabilities of variational autoencoders (VAEs) by employing energy-based priors. Gilhan Kim and Daniel K. Park at Yonsei University present Boltzmann-machine, prior VAEs (BM-VAEs) trained using quantum annealing, utilising three distinct operational modes within a single system to improve performance. The research shows stable and efficient training across multiple datasets using up to 2000 qubits on a D-Wave Advantage2 processor, achieving faster convergence and reduced reconstruction loss compared to standard Gaussian-prior VAEs. It sharply advances the field by enabling both unconditional generation, a feature absent in plain autoencoders, and conditional generation through latent biasing, using learned interactions between latent variables

Quantum annealing enables substantial performance gains in Boltzmann-machine prior variational autoencoders

A 35% reduction in reconstruction loss, previously unattainable with standard autoencoders, has been achieved. This improvement stems from the development of Boltzmann-machine, prior VAEs (BM-VAEs) trained using quantum annealing, which facilitates the learning of structured interactions between latent variables. These latent variables represent the hidden features the model extracts from data, effectively forming a compressed and abstract representation of the input. The system utilises diabatic, slower, and conditional quantum annealing, three distinct operational modes, within a single generative system, optimising both training and data generation. Diabatic annealing provides a rapid initial exploration of the energy landscape, while slower annealing refines the solution with increased precision. Conditional annealing then focuses on specific regions of the latent space, enabling targeted data generation. This multi-modal approach is crucial for navigating the complex energy landscapes inherent in Boltzmann machines

D-Wave scientists established stable and efficient training across multiple datasets using up to 2000 qubits on a D-Wave Advantage2 processor. This unlocked both unconditional and conditional generation capabilities, features not found in plain autoencoders. Experiments on the CelebA dataset demonstrated superior performance compared to standard autoencoders, as the learned Boltzmann prior allows direct sampling from the latent distribution. The CelebA dataset, comprising over 200,000 celebrity face images, provided a robust benchmark for evaluating the generative capabilities of the BM-VAE. However, current results are limited to controlled laboratory conditions and do not yet demonstrate robust performance with noisy or incomplete real-world data, representing a key challenge for practical implementation. The susceptibility to noise stems from the delicate balance required within the quantum annealing process, where even minor disturbances can affect the final solution.

The technique proved crucial because standard variational autoencoders (VAEs) often struggle to learn intricate relationships between data characteristics. VAEs create a compressed ‘understanding of data, akin to a detailed book summary, but frequently assume these characteristics are independent. This simplification, while computationally convenient, limits their ability to model complex, real-world phenomena where features are often highly correlated. To overcome these limitations, a Boltzmann machine prior was employed, a type of energy-based model where connections between latent variables are unrestricted. This allows for the creation of more complex and nuanced data representations, moving beyond the simplified assumptions of traditional VAEs. The energy-based framework defines a probability distribution over the latent space based on the ‘energy’ of each configuration, with lower energy states representing more probable configurations. The quantum approach was favoured as it natively implements general Ising Hamiltonians, enabling sampling from complex Boltzmann machines without structural constraints, a task that challenges classical methods. Classical methods often require approximations or simplifications to make the sampling process tractable, potentially sacrificing accuracy.

Boltzmann machine prior variational autoencoders trained via quantum annealing

Generative models, systems that learn to create new data resembling their training set, are receiving increasing attention from researchers tackling complex problems in fields such as drug discovery and materials science. In drug discovery, generative models can propose novel molecular structures with desired properties, while in materials science, they can design new materials with specific characteristics. Standard variational autoencoders, a popular type of generative model, often struggle to capture intricate relationships within data, relying on simplified assumptions about how different characteristics interact. This new work offers a compelling solution by employing quantum annealing, a specialised computing technique, to train more expressive energy-based priors, allowing the model to learn more subtle connections. Quantum annealing leverages quantum-mechanical effects to efficiently search for the lowest energy state of a system, which corresponds to the most probable configuration of the latent variables.

Although quantum annealing requires specialised hardware and is not yet universally accessible, this work demonstrates a valuable pathway for enhancing generative models. The team successfully trained these models, termed Boltzmann-machine, prior VAEs, using up to 2000 qubits, achieving faster convergence and more detailed data recreation than traditional methods. The convergence rate, a measure of how quickly the model learns, was significantly improved, reducing training time and computational cost. The ability to generate new data based on specific attributes offers significant potential across diverse applications, justifying further investigation into this hybrid quantum-classical approach. For example, in image generation, users could specify desired features such as colour, shape, or texture, and the model would generate images that match those criteria.

This research establishes a novel method for training generative models by integrating quantum annealing with variational autoencoders, creating Boltzmann-machine prior VAEs. By using a D-Wave processor, scientists at D-Wave established stable training and improved data recreation compared to conventional systems, achieved through the three operational modes, each optimising a specific stage of the generative process. The system surpasses standard autoencoders by enabling unconditional data generation, creating new data without specific prompts. Furthermore, it also allows for targeted attribute control during generation. This combination of capabilities positions the new approach as a promising avenue for advanced generative modelling. The ability to manipulate the latent space allows for fine-grained control over the generated output, opening up possibilities for creative applications and scientific exploration. Further research will focus on improving the robustness of the system to noisy data and scaling the approach to even larger and more complex datasets.

The research successfully trained Boltzmann-machine, prior VAEs using up to 2000 qubits on a D-Wave Advantage2 processor. This new approach improves generative models by enabling unconditional data generation and targeted control over attributes, capabilities not found in standard autoencoders. Training with quantum annealing resulted in faster convergence and lower reconstruction loss compared to a Gaussian-prior VAE. The authors intend to improve the system’s robustness to noisy data and scale the approach to larger datasets.