[論文レビュー] Latent Style-based Quantum GAN for high-quality Image Generation

Su Yeon Chang, Supanut Thanasilp+3 more|arXiv (Cornell University)|Jun 4, 2024

Computational Physics and Python Applications被引用数 5

ひとこと要約

tldr: LaSt-QGANは古典的オートエンコーダを用いて画像を潜在空間に写像し、量子ジェネレータが潜在特徴を生成することで、MNIST、FashionMNIST、および SAT4 に対する古典GANと競合する性能を持つ大規模画像生成を実現する。さらにショットノイズとボaran plateau（barren plateau）を分析。

ABSTRACT

Quantum generative modeling is among the promising candidates for achieving a practical advantage in data analysis. Nevertheless, one key challenge is to generate large-size images comparable to those generated by their classical counterparts. In this work, we take an initial step in this direction and introduce the Latent Style-based Quantum GAN (LaSt-QGAN), which employs a hybrid classical-quantum approach in training Generative Adversarial Networks (GANs) for arbitrary complex data generation. This novel approach relies on powerful classical auto-encoders to map a high-dimensional original image dataset into a latent representation. The hybrid classical-quantum GAN operates in this latent space to generate an arbitrary number of fake features, which are then passed back to the auto-encoder to reconstruct the original data. Our LaSt-QGAN can be successfully trained on realistic computer vision datasets beyond the standard MNIST, namely Fashion MNIST (fashion products) and SAT4 (Earth Observation images) with 10 qubits, resulting in a comparable performance (and even better in some metrics) with the classical GANs. Moreover, we analyze the barren plateau phenomena within this context of the continuous quantum generative model using a polynomial depth circuit and propose a method to mitigate the detrimental effect during the training of deep-depth networks. Through empirical experiments and theoretical analysis, we demonstrate the potential of LaSt-QGAN for the practical usage in the context of image generation and open the possibility of applying it to a larger dataset in the future.

研究の動機と目的

• Motivate and develop a hybrid classical-quantum GAN (LaSt-QGAN) capable of generating large-size images.
• Leverage a pre-trained convolutional autoencoder to map high-dimensional images into a latent space for efficient quantum generation.
• Train a quantum generator with a classical discriminator to reproduce latent features and reconstruct images via the autoencoder.
• Evaluate LaSt-QGAN on MNIST, FashionMNIST, and SAT4 and compare with a matched classical GAN.
• Investigate robustness to shot noise and analyze barren plateau phenomena to inform trainability of continuous quantum generative models.

提案手法

• Use a pre-trained convolutional autoencoder to encode images into a latent space of dimension Dℓ; train a quantum generator Gθ in this latent space and a classical discriminator Dφ with Wasserstein loss and gradient penalty.
• Employ a parameterized quantum circuit (style-based generator) where latent noise z is embedded into rotation angles; use L layers with θℓ = Wℓ z + bℓ (data reuploading concept).
• Measure expectation values ⟨σx⟩ and ⟨σz⟩ on n qubits as latent features and concatenate to form a 2n-dimensional feature vector for the discriminator.
• Reconstruct images by passing generated latent features through the (pre-trained) autoencoder’s decoder; train with a Wasserstein distance objective to match real latent features to fake ones.
• Compare multiple quantum circuit architectures (Circuits 1–3) and quantify performance with FID, IS, and JSD on both features and reconstructed images.
• Assess training dynamics and robustness to shot noise, and analyze barren plateau behavior to propose initialization strategies for polynomial-depth circuits.

(a) Discrete Quantum GAN

実験結果

リサーチクエスチョン

• RQ1 Can LaSt-QGAN generate large-size images by operating in a latent space mapped from high-dimensional data?
• RQ2 How does LaSt-QGAN performance compare to a classical GAN with a similar parameter count across MNIST, FashionMNIST, and SAT4 datasets?
• RQ3 What is the impact of quantum circuit depth and architecture on generation quality and training stability?
• RQ4 How robust is LaSt-QGAN to shot-noise (finite sampling) and what training strategies mitigate potential barren plateau effects?

主な発見

G_theta config	N_Θ	FID ↓	IS ↑	JSD (features/ 10^-2) ↓	JSD (images/ 10^-2) ↓
Circ. 1 ( d=2 )	1360	17.2±0.35	8.29±0.02	0.79±0.05	1.63±0.09
Circ. 1 ( d=4 )	2280	14.85±0.34	8.49±0.04	0.75±0.07	1.49±0.18
Circ. 1 ( d=6 )	3200	14.13±0.73	8.53±0.05	0.71±0.07	1.29±0.10
Circ. 2 ( d=2 )	1010	19.13±0.54	8.10±0.06	1.22±0.19	2.08±0.17
Circ. 2 ( d=4 )	1690	16.2±0.32	8.34±0.03	0.94±0.09	1.66±0.17
Circ. 2 ( d=6 )	2370	14.85±0.61	8.47±0.06	0.85±0.05	1.39±0.11
Circ. 3 ( d=2 )	3300	14.29±0.38	8.50±0.04	0.76±0.06	1.50±0.12
Circ. 3 ( d=4 )	6600	12.72±0.40	8.65±0.05	0.71±0.07	1.14±0.12
Circ. 3 ( d=6 )	9900	11.99±0.56	8.71±0.04	0.72±0.09	1.13±0.12
Classical [50,30]	2960	18.24±3.6	8.24±0.28	3.74±1.64	4.51±2.0
Classical [100,50]	7660	12.56±0.91	8.80±0.06	1.18±0.17	1.56±0.13

• LaSt-QGAN can generate large-size images and achieves competitive or superior metrics (FID, IS, JSD) compared with a classical GAN of similar size across MNIST, FashionMNIST, and SAT4.
• Faster convergence and higher stability are observed for LaSt-QGAN than the classical GAN on MNIST and FashionMNIST across several circuit depths.
• For MNIST and FashionMNIST, LaSt-QGAN attains lower JSD values and favorable FID/IS trends compared to the classical counterpart, indicating better learning of data distribution and diversity.
• On SAT4, LaSt-QGAN outperforms the classical GAN on all evaluated metrics while using roughly half the number of parameters.
• t-SNE visualization shows generated features form class-separated clusters, suggesting preserved latent structure in generation.
• The study provides a method to mitigate barren plateau effects at initialization with small-angle starts for polynomial-depth circuits, enhancing trainability for continuous quantum generative models.

(b) Continuous Quantum GAN