Quantum Computers Harness Randomness to Build More Efficient Models

Arunava Majumder and colleagues at University of Innsbruck present a key advance in variational measurement-based quantum computation (MBQC), a technique utilising entangled quantum states and single-qubit measurements to process information. Their research addresses a key limitation of existing variational MBQC approaches, which often require a substantial increase in trainable parameters, scaling with both the width and depth of the model, potentially hindering optimisation and model training. By introducing a restricted VMBQC model with only a single additional trainable parameter, they successfully extend the capabilities of unitary models to channel-based ones. This minimal extension enables the generation of probability distributions beyond the reach of conventional unitary models, opening new avenues for generative modelling with quantum computers.

Single parameter control unlocks enhanced quantum generative modelling

The scaling of trainable parameters in variational measurement-based quantum computation (VMBQC) has been dramatically reduced. Previously, VMBQC required N × D parameters, but the new restricted model achieves comparable performance with only one. This represents a key leap, as the prior parameter scaling hindered optimisation and model training, particularly as the width (N) and depth (D) of the model increased. This single parameter is sufficient to generate probability distributions inaccessible to unitary models, expanding the potential of quantum generative modelling.

A restricted variational measurement-based quantum computation (VMBQC) model surpasses the capabilities of its unitary counterparts with just one trainable parameter. This contrasts sharply with earlier VMBQC models, which scaled parameter requirements by a factor of N × D, where N represents the number of logical qubits and D denotes the model’s depth. Algebraic and numerical analysis demonstrated that this single parameter allows the VMBQC model to generate probability distributions impossible for corresponding unitary models to replicate, broadening the range of accessible quantum generative modelling. The framework utilises measurement-based quantum computation, employing a highly entangled resource state and single-qubit measurements to perform calculations. Current results focus on relatively small systems, and scalability to the complex datasets needed for real-world applications remains to be demonstrated.

Measurement-based quantum computation fundamentally differs from the more commonly known circuit model of quantum computing. Instead of manipulating qubits through a sequence of gates, MBQC relies on preparing a highly entangled multi-qubit state, known as a resource state, such as a cluster state or graph state. Computation then proceeds by performing single-qubit measurements on this resource state, with the measurement outcomes determining the subsequent measurements. This approach offers potential advantages in terms of fault tolerance and hardware requirements. However, the inherent randomness introduced by quantum measurements necessitates careful consideration when implementing computational tasks. In standard MBQC, classical feedback is used to correct for this randomness and ensure deterministic unitary computation. Variational MBQC, however, deliberately embraces this indeterminacy.

The significance of reducing the number of trainable parameters lies in the challenges associated with optimising quantum models. Quantum computers are still in their early stages of development, and the number of qubits available is limited. Furthermore, the process of optimising a quantum circuit or model is often computationally expensive and prone to errors. Many parameters exacerbate these difficulties, making it harder to find the optimal settings for the model. By reducing the parameter count to a single value, the researchers have significantly simplified the optimisation landscape, potentially enabling the training of more complex and powerful generative models on near-term quantum hardware.

Single parameter extension unlocks scalable quantum generative modelling

Variational measurement-based quantum computation, or VMBQC, offers a promising route to generative modelling by harnessing the inherent randomness of quantum measurement. Instead of correcting this randomness, VMBQC exploits it to create more complex probability distributions. Earlier VMBQC models suffered from a significant drawback, requiring a number of adjustable parameters that grew rapidly with the size and complexity of the quantum system, scaling with both width and depth.

The increased complexity of these earlier variational measurement-based quantum computation models presented a significant hurdle to practical application. This new model achieves comparable performance with a dramatically simplified structure, needing only a single additional adjustable parameter. This minimal extension allows the creation of probability distributions beyond the reach of standard quantum computers, unlocking new generative modelling capabilities.

A simplified quantum model for generative modelling has been created, reducing adjustable parameters to a single addition. Variational measurement-based quantum computation, or VMBQC, builds complex probability distributions from the randomness of quantum measurement. Previous models required parameters that increased rapidly with system size, hindering practical use. Deliberately utilising the inherent randomness of quantum measurement, scientists created a model extending standard quantum settings with only one additional parameter. This restricted model generates probability distributions that comparable, traditional quantum systems cannot replicate, expanding the possibilities for quantum generative modelling.

The core innovation lies in the ability to generate non-unitary probability distributions. Unitary transformations, which preserve the norm of quantum states, are the foundation of many quantum algorithms. However, they are limited in their ability to produce certain types of probability distributions. Non-unitary operations, which allow for the loss of probability amplitude, are necessary to generate more general distributions. The researchers demonstrate that their restricted VMBQC model can effectively implement such non-unitary transformations using the single additional parameter, effectively extending the expressivity of the quantum model. This is achieved by carefully controlling the influence of the measurement outcomes on the subsequent measurements within the MBQC framework.

The implications of this work extend to various areas of quantum machine learning. Generative models are used for a wide range of tasks, including image generation, natural language processing, and drug discovery. A quantum generative model capable of producing more complex and realistic distributions could potentially outperform classical models in these areas. Furthermore, the reduced parameter count makes the model more amenable to implementation on near-term quantum devices, paving the way for practical applications soon. While the current demonstration focuses on relatively small systems, the researchers suggest that the approach is scalable and could be extended to larger and more complex datasets. Future work will focus on exploring the full potential of this restricted VMBQC model and demonstrating its performance on real-world generative modelling tasks, as well as investigating methods to further optimise the single parameter for specific applications.

The researchers developed a new quantum computing model that generates probability distributions beyond the capabilities of traditional systems. This was achieved by extending a measurement-based quantum computation framework with only one additional adjustable parameter. This minimal extension allows the model to create more complex distributions, which is important for improving quantum generative modelling. The authors intend to explore this model’s potential with real-world data and optimise the single parameter for specific tasks.