A data-driven and model-based accelerated Hamiltonian Monte Carlo method for Bayesian elliptic inverse problems

Sijing Li, Cheng Zhang, Zhiwen Zhang, Hongkai ZhaoStat. Comput 2021 · OpenReview

TL;DR

This paper introduces a novel data-driven and model-based approach to accelerate Hamiltonian Monte Carlo (HMC) for Bayesian inverse problems governed by elliptic partial differential equations (PDEs). By leveraging the intrinsic low-dimensional structure of these PDEs and constructing data-driven bases, the method significantly reduces computational costs while maintaining accuracy and exploration efficiency.

What Problem Does This Paper Address?

Bayesian inverse problems are crucial for inferring unknown parameters from noisy observations, often modeled by complex systems like elliptic partial differential equations (PDEs). While Hamiltonian Monte Carlo (HMC) is a powerful tool for exploring the posterior distribution in these problems, it suffers from a significant computational bottleneck: repeatedly solving the PDE and its derivatives to evaluate the Hamiltonian. This becomes prohibitively expensive for large-scale problems, hindering their practical application. The paper aims to overcome this by developing a method that drastically reduces the computational cost of HMC without sacrificing its sampling efficiency.

Methods

The core of the proposed method lies in a two-pronged strategy:

  • Data-Driven Basis Construction: The authors exploit the inherent low-dimensional structure of elliptic PDEs. They first collect solution samples (either from real measurements or during an initial HMC "burn-in" phase) and then use Proper Orthogonal Decomposition (POD) to extract a set of data-driven basis functions. These bases capture the essential modes of the solution space, enabling significant dimension reduction.
  • Model-Based Parameter-to-Solution Map: Once the data-driven bases are established, the authors train neural networks to approximate the mapping from the input parameters (which define the PDE's coefficients) to the coefficients of the solution in the data-driven basis. This "parameter-to-solution map" allows for rapid computation of the solution and its derivatives for new parameter samples, bypassing the need to solve the full PDE each time.

This approach is integrated into the HMC framework. After an initial burn-in phase to gather data for training, the accelerated HMC uses the learned parameter-to-solution maps to efficiently compute the necessary quantities for proposing new samples, thereby accelerating the overall sampling process.

Datasets

The paper does not introduce new benchmark datasets. Instead, it generates synthetic data by solving the elliptic PDE (1) with various realizations of the random coefficient. These solutions and their derivatives are used for:

  1. Constructing the data-driven basis functions via POD.
  2. Training the neural networks for the parameter-to-solution maps.

The numerical experiments are conducted on a 2D elliptic PDE with a log-normal coefficient, first with isotropic and then with anisotropic heterogeneity.

Key Insights and Ideas

  • Intrinsic Low-Dimensional Structure: Elliptic PDEs, even with complex random coefficients, often possess an underlying low-dimensional structure in their solution space. This structure can be effectively captured by data-driven bases.
  • Synergy of Data and Model: The proposed method beautifully combines data-driven techniques (POD for basis extraction) with model-based insights (understanding the PDE's structure and the parameter-to-solution relationship). This synergy is key to its efficiency.
  • Parameter-to-Solution Maps as Accelerators: Learning a map from input parameters to the solution's representation in a reduced basis is a powerful way to avoid repeated expensive PDE solves. Neural networks prove to be effective function approximators for these maps.
  • Balancing Efficiency and Exploration: The method achieves a good balance between computational efficiency (due to dimension reduction and fast approximations) and exploration efficiency (maintaining high acceptance rates in HMC).

Key Results

The accelerated HMC method demonstrates significant improvements in computational efficiency compared to standard HMC and a random network surrogate method.

  • Speedup: Numerical experiments show a substantial reduction in computation time, with the accelerated HMC being orders of magnitude faster in some cases. For instance, Table 1 shows significantly lower "s/Iter" (seconds per iteration) for the data-driven method.
  • Accuracy: The method maintains high accuracy, with relative errors for posterior means and standard deviations being very small (e.g., 0.047 and 0.024 respectively in the second example).
  • Acceptance Rate: The acceptance rate remains high and stable, even as the dimensionality of the input parameters increases, indicating good exploration of the posterior.
  • Effective Sample Size: The time-normalized effective sample size (ESS/s) is considerably higher for the proposed method, indicating more efficient sampling per unit of time.

Limitations and Future Work

  • Burn-in Stage: The method still requires an initial burn-in stage of standard HMC to collect data for training. While this is a one-time cost, it can still be significant for very complex problems.
  • Data Generation: The quality and quantity of training data are crucial for constructing accurate data-driven bases and parameter-to-solution maps. Generating sufficient and representative data can be challenging, especially for high-dimensional problems.
  • Generalizability: While the paper focuses on elliptic PDEs, extending this approach to other types of PDEs or inverse problems might require further investigation.

The authors suggest that their approach strikes a good balance and achieves significant savings, implying that further applications and extensions are promising. The numerical results demonstrate the effectiveness of their data-driven basis and neural network approximation for accelerating Hamiltonian dynamics.

Summary updated

Abstract

In this paper, we consider a Bayesian inverse problem modeled by elliptic partial differential equations (PDEs). Specifically, we propose a data-driven and model-based approach to accelerate the Hamiltonian Monte Carlo (HMC) method in solving large-scale Bayesian inverse problems. The key idea is to exploit (model-based) and construct (data-based) intrinsic approximate low-dimensional structure of the underlying problem which consists of two components—a training component that computes a set of data-driven basis to achieve significant dimension reduction in the solution space, and a fast solving component that computes the solution and its derivatives for a newly sampled elliptic PDE with the constructed data-driven basis. Hence we develop an effective data and model-based approach for the Bayesian inverse problem and overcome the typical computational bottleneck of HMC—repeated evaluation of the Hamiltonian involving the solution (and its derivatives) modeled by a complex system, a multiscale elliptic PDE in our case. Finally, we present numerical examples to demonstrate the accuracy and efficiency of the proposed method.
BibTeX
@inproceedings{li2021data,
  author = {Sijing Li and Cheng Zhang and Zhiwen Zhang and Hongkai Zhao},
  title = {{A data-driven and model-based accelerated Hamiltonian Monte Carlo method for Bayesian elliptic inverse problems}},
  booktitle = {Stat. Comput},
  year = {2021},
  url = {https://openreview.net/forum?id=LsFclD_qaHH},
}
Figure