Quantum MeanFlow: single-shot generative sampling on NISQ hardware
Quantum MeanFlow enables single-step MNIST image generation on NISQ hardware with reduced sampling costs.
Key Findings
Methodology
Quantum MeanFlow (QMF) learns average velocity fields over time intervals to enable single-step sampling, avoiding the sequential evaluations of Quantum Flow Matching (QFM). It uses parameterized quantum circuits combined with classical autoencoders to process MNIST data.
Key Results
- QMF achieves 63% accuracy in simulation for single-step MNIST generation, boosted to 100% via BoN sampling.
- Hardware single-step accuracy is 20%, improved to 60% with BoN sampling.
- QMF single-step FID is 28.26, lower than QFM's 10.62 but significantly better than the pre-training value of 150.3.
Significance
QMF reduces sampling costs on quantum hardware, making it suitable for current NISQ devices. It opens new avenues for applying quantum generative models to classical data.
Technical Contribution
Introduced a quantum generative model based on average velocity fields, optimized circuit design for NISQ hardware, and employed BoN sampling to mitigate hardware noise.
Novelty
QMF is the first adaptation of MeanFlow to quantum generative models, demonstrating feasibility for classical data generation.
Limitations
- QMF produces lower-quality images compared to QFM, with higher FID scores.
- Hardware noise significantly affects sampling quality, especially at low shot counts.
- Training still relies on classical simulation, limiting full quantum implementation.
Future Work
Future work could focus on optimizing circuit designs to reduce noise and testing QMF on more complex datasets.
AI Executive Summary
Quantum generative models offer new possibilities for machine learning, particularly in generating complex data distributions. However, existing methods like Quantum Flow Matching (QFM) require multi-step sampling, leading to high costs and inefficiencies on quantum hardware.
This paper introduces Quantum MeanFlow (QMF), which learns average velocity fields to enable single-step sampling. Compared to QFM, QMF significantly reduces quantum circuit evaluations, making it more suitable for NISQ hardware. Experiments on the MNIST dataset show that QMF achieves 63% accuracy in simulation, boosted to 100% via BoN sampling.
While QMF sacrifices image quality (FID 28.26 vs. QFM's 10.62), its efficiency advantage makes it a significant step forward for quantum generative models. Future research could optimize circuit designs and explore applications on more complex datasets.
Deep Analysis
Background
Quantum computing has gained attention in recent years, especially in quantum machine learning. Classical generative models like GANs and VAEs have achieved success, but quantum generative models provide new ways to explore complex data distributions. Existing Quantum Flow Matching methods are effective but suffer from high costs due to multi-step sampling.
Core Problem
Applying quantum generative models to classical data remains challenging due to low sampling efficiency and hardware noise. Designing efficient quantum models for NISQ hardware is a critical bottleneck.
Innovation
QMF enables single-step sampling by learning average velocity fields, reducing quantum circuit evaluation costs. Innovations include: 1) extending MeanFlow to quantum models; 2) optimizing quantum circuit designs to reduce noise; 3) employing BoN sampling to improve hardware-generated image quality.
Methodology
- �� Classical autoencoder processes MNIST images into 32-dimensional latent vectors
- �� Parameterized quantum circuit encodes latent vectors and class labels using angle encoding
- �� QMF circuit design includes 24 layers of rotation gates and CNOT chains, optimized for IBM Heron devices
- �� Training uses classical simulation with AdamW optimizer and MF loss function
- �� Sampling generates images in a single step using learned average velocity fields
Experiments
Experiments use the MNIST dataset, with images downscaled to 16×16. QMF and QFM models are trained on classical simulators, and sampling is performed on IBM Heron hardware. Results compare sampling quality across shot counts and evaluate BoN strategies.
Results
QMF achieves 63% accuracy in simulation, 20% on hardware, and 60% with BoN sampling. Single-step FID is 28.26, significantly improved from the pre-training value of 150.3.
Applications
QMF is suitable for generating handwritten digit images and can be used for quantum generative model research and hardware performance evaluation.
Limitations & Outlook
QMF produces lower-quality images, and hardware noise significantly impacts sampling performance. Training relies on classical simulation, limiting full quantum implementation.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen. Traditional methods require step-by-step preparation, like chopping vegetables and boiling soup, where each step depends on the previous one. This is like Quantum Flow Matching, which requires multi-step sampling. Quantum MeanFlow is like using a high-powered blender to process everything at once, saving time but slightly compromising the dish's texture.
ELI14 Explained like you're 14
Hey, imagine you're playing a video game! In normal mode, you have to complete tasks step by step, like fighting monsters and upgrading gear. That's like Quantum Flow Matching. But QMF is like turning on 'fast mode,' skipping straight to the final boss. Sure, the graphics might be less detailed, but you save time and can play more rounds!
Glossary
Quantum Flow Matching
A quantum generative method that generates target distributions via multi-step sampling.
Used for generating MNIST images in prior work.
Quantum MeanFlow
A quantum generative method enabling single-step sampling by learning average velocity fields.
Proposed in this paper to reduce sampling costs.
BoN Sampling
A rejection sampling strategy that selects the best sample to improve quality.
Used to mitigate hardware noise effects.
FID (Fréchet Distance)
A metric for evaluating the quality of generated images by comparing distributions.
Used to assess image quality in this study.
NISQ Hardware
Noisy Intermediate-Scale Quantum devices available today.
IBM Heron devices used in this paper's experiments.
Open Questions Unanswered questions from this research
- 1 How can quantum circuit designs be optimized to further reduce noise?
- 2 Can QMF be applied to more complex datasets with high-quality results?
Applications
Immediate Applications
Quantum Generative Model Research
Provides tools for exploring classical data generation using quantum methods.
Hardware Performance Evaluation
Tests NISQ hardware efficiency and noise impact using QMF.
Long-term Vision
Complex Data Generation
Future applications could include generating high-resolution images or complex data distributions.
Abstract
Quantum generative models offer a promising framework for exploring whether quantum computation can enhance generative machine learning. Flow matching is a generative method in which samples are generated by transporting a simple, known distribution to the target data distribution with a learned velocity field. Its quantum counterpart, known as quantum flow matching (QFM), was introduced recently, and, like its classical counterpart, requires integrating an ordinary differential equation over many time steps during inference. As each step requires the output from the previous step, the circuit submission is sequential and a drawback on quantum computers as they have high input/output costs. To alleviate this problem, we introduce Quantum MeanFlow (QMF), the quantum analogue of the MeanFlow formulation, which allows single-step sample generation. While the QFM learns an instantaneous velocity field at each time step, QMF learns the average velocity over a time interval. We use a parameterized quantum circuit to learn these velocity fields and benchmark the two methods on the MNIST dataset. We show that while single-step QMF has lower image quality compared to multi-step QFM, it performs better than the single-step QFM sampling at every shot count. Both of our models are executed on IBM quantum computers and best-of-N rejection sampling recovers most of the accuracy lost to device noise without modifying the circuit. This is especially advantageous for QMF which has only one circuit evaluation per image. Here, We establish QMF as a viable method for single-step quantum generative sampling, saving on quantum circuit evaluations per generated sample.