Quantum-Boosted High-Fidelity Deep Learning

Feng-ao Wang, Shaobo Chen, Yao Xuan, Junwei Liu, Qi Gao, Hongdong Zhu, Junjie Hou, Lixin Yuan, Jinyu Cheng, Chenxin Yi, Hai Wei, Yin Ma, Tao Xu, Kai Wen, Yixue LiarXiv 2025 · DOI · arXiv

Introduction

Variational Autoencoders (VAEs) are fundamental tools in modern probabilistic deep learning, widely used for dimensionality reduction and latent feature extraction. However, a significant bottleneck in their performance is the near-universal reliance on Gaussian priors. While mathematically convenient for optimization, the assumption that latent features follow a simple normal distribution is often too rigid for real-world datasets, such as those found in biological systems, which exhibit complex, non-Gaussian "energy landscapes."

The Boltzmann distribution, rooted in statistical physics, offers a far more expressive alternative. However, calculating its partition function is computationally intractable on classical hardware due to the exponential growth of the state space. This paper presents the Quantum Boltzmann Machine-Variational Autoencoder (QBM-VAE), a hybrid framework that leverages quantum sampling to bridge this gap, enabling high-fidelity representation of million-scale biological datasets.

Methodology: Hybrid Quantum-Classical Architecture

The QBM-VAE architecture splits the computational load between classical and quantum systems. A classical computer manages the neural network's encoder and decoder, while a Coherent Ising Machine (CIM)—a specialized quantum processor—performs Boltzmann sampling.

Overview of the QBM-VAE framework, comparing priors and detailing the hybrid quantum-classical architecture.

The model maps high-dimensional input data into a discrete, binarized latent space. By utilizing the CIM to solve the Ising model, the framework performs large-scale Boltzmann sampling, allowing for the optimization of the Kullback-Leibler (KL) divergence between the data's posterior and a rigorous Boltzmann prior. A key technical achievement is the implementation of a reparameterization technique that permits gradient-based optimization (backpropagation) through this discrete latent space. Furthermore, the CIM hardware demonstrated the necessary stability for deep learning, operating continuously for over 12 hours with thousands of fully connected spins.

Key Results and Insights

The QBM-VAE was evaluated on several massive single-cell RNA-sequencing (scRNA-seq) datasets, including the Human Lung Cell Atlas, totaling over one million cells. These datasets represent a paradigmatic challenge for traditional VAEs due to their high-dimensional, non-Gaussian complexity.

The model demonstrated several key advantages over state-of-the-art classical methods like scVI:

  • Enhanced Biological Conservation: QBM-VAE consistently outperformed existing methods in both biological signal conservation and batch effect correction.
  • Rare Cell Resolution: A significant breakthrough was the model's ability to uniquely resolve rare cell populations, such as macrophages and mast cells, into distinct clusters—a task where classical models often fail by intermingling them with common types.
  • Topological Preservation: The Boltzmann-informed latent space achieved a 20-30% improvement in the scGraph score, indicating that it better recapitulates the intrinsic structural relationships of biological data.
Downstream applications of QBM-VAE in cell classification and trajectory inference.

In downstream tasks, QBM-VAE achieved superior accuracy in cell-type classification and a near-perfect trajectory conservation score of 0.995. These results suggest that embedding a "physical inductive bias" into the model allows it to better "understand" the hierarchical organization and energy-driven dynamics of biological systems.

Conclusion

The QBM-VAE framework demonstrates a practical path toward quantum advantage in deep learning. By replacing oversimplified Gaussian assumptions with physically-grounded Boltzmann distributions, the model achieves superior fidelity in representing complex data. Future research will explore the application of this framework to protein design and materials science, as well as its potential integration into generative diffusion models.

Summary updated

BibTeX
@misc{wang2025quantum,
  author = {Feng-ao Wang and Shaobo Chen and Yao Xuan and Junwei Liu and Qi Gao and Hongdong Zhu and Junjie Hou and Lixin Yuan and Jinyu Cheng and Chenxin Yi and Hai Wei and Yin Ma and Tao Xu and Kai Wen and Yixue Li},
  title = {{Quantum-Boosted High-Fidelity Deep Learning}},
  year = {2025},
  eprint = {2508.11190},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.11190},
}
Figure