Truncated proposals for scalable and hassle-free simulation-based inference

Michael Deistler, Pedro J. Gonçalves, Jakob H. Macke

Abstract

Simulation-based inference (SBI) solves statistical inverse problems by repeatedly running a stochastic simulator and inferring posterior distributions from model-simulations. To improve simulation efficiency, several inference methods take a sequential approach and iteratively adapt the proposal distributions from which model simulations are generated. However, many of these sequential methods are difficult to use in practice, both because the resulting optimisation problems can be challenging and efficient diagnostic tools are lacking. To overcome these issues, we present Truncated Sequential Neural Posterior Estimation (TSNPE). TSNPE performs sequential inference with truncated proposals, sidestepping the optimisation issues of alternative approaches. In addition, TSNPE allows to efficiently perform coverage tests that can scale to complex models with many parameters. We demonstrate that TSNPE performs on par with previous methods on established benchmark tasks. We then apply TSNPE to two challenging problems from neuroscience and show that TSNPE can successfully obtain the posterior distributions, whereas previous methods fail. Overall, our results demonstrate that TSNPE is an efficient, accurate, and robust inference method that can scale to challenging scientific models.

Introduction

In many scientific fields, researchers rely on complex simulators (e.g., for climate modeling, neuroscience, or particle physics) to describe observed phenomena. While these simulators can generate data, they often lack an explicit "likelihood function," making standard Bayesian inference impossible. Simulation-Based Inference (SBI) addresses this by training neural networks to learn the relationship between parameters and simulated data. To increase efficiency, Sequential methods iteratively refine parameter sampling, focusing on regions that resemble observed data. However, existing state-of-the-art methods like Automatic Posterior Transformation (APT) often suffer from stability issues and "leakage," where the model predicts posterior mass in regions the prior deems impossible.

Methods: Truncated Sequential Neural Posterior Estimation (TSNPE)

The authors introduce Truncated Sequential Neural Posterior Estimation (TSNPE) to solve these robustness and scalability issues. The core innovation is a deceptively simple strategy: Truncation.

The Truncated Proposal

Instead of using complex loss functions to correct for biased proposals, TSNPE samples from the prior but rejects any parameters falling outside the current estimated posterior’s Highest-Probability Region (HPR). Because the proposal is simply a truncated version of the prior, the authors prove that standard maximum-likelihood training can be used in every round, eliminating the optimization instabilities found in APT.

The neural density estimator architecture used within the TSNPE framework to learn the posterior distribution.

Scalable Diagnostics

TSNPE leverages normalizing flows for fast density evaluation and sampling. This enables Simulation-Based Coverage Calibration (SBCC), a tool that allows researchers to verify if the model is well-calibrated (e.g., checking if 95% confidence intervals actually contain the truth 95% of the time) after every training round, without requiring expensive MCMC steps.

Key Results

TSNPE was evaluated against standard SBI benchmarks and two high-dimensional neuroscience models: a 31-parameter Pyloric Network model and a 20-parameter L5PC neuron model.

Overcoming the "Leakage" Wall

In high-dimensional problems, APT's leakage becomes a total failure mode. For the Pyloric network, APT's rejection rate for valid samples exceeded 99.9999%, causing it to fail after just three rounds. In contrast, TSNPE successfully identified posterior distributions matching experimental data.

Comparison of posterior estimates. APT exhibits a "leakage" artifact (central spike), whereas TSNPE accurately recovers the ground truth modes.

High-Dimensional Scaling

On the L5PC neuron model, TSNPE produced the first known successful posterior for this landmark problem. This was achieved despite the fact that 99.98% of prior samples produced invalid data, demonstrating the method's extreme robustness in difficult parameter spaces.

Conclusion

By shifting complexity from the loss function to the sampling strategy, TSNPE provides a robust, "hassle-free" framework for simulation-based inference. It maintains parity with existing methods on simple tasks while succeeding on complex, high-dimensional scientific models where previous state-of-the-art methods fail. The inclusion of scalable diagnostics like SBCC further ensures that these models remain reliable for high-stakes scientific applications.

Summary updated

Abstract

Simulation-based inference (SBI) solves statistical inverse problems by repeatedly running a stochastic simulator and inferring posterior distributions from model-simulations. To improve simulation efficiency, several inference methods take a sequential approach and iteratively adapt the proposal distributions from which model simulations are generated. However, many of these sequential methods are difficult to use in practice, both because the resulting optimisation problems can be challenging and efficient diagnostic tools are lacking. To overcome these issues, we present Truncated Sequential Neural Posterior Estimation (TSNPE). TSNPE performs sequential inference with truncated proposals, sidestepping the optimisation issues of alternative approaches. In addition, TSNPE allows to efficiently perform coverage tests that can scale to complex models with many parameters. We demonstrate that TSNPE performs on par with previous methods on established benchmark tasks. We then apply TSNPE to two challenging problems from neuroscience and show that TSNPE can successfully obtain the posterior distributions, whereas previous methods fail. Overall, our results demonstrate that TSNPE is an efficient, accurate, and robust inference method that can scale to challenging scientific models.
BibTeX
@inproceedings{deistler2022truncated,
  author = {Michael Deistler and Pedro J. Gonçalves and Jakob H. Macke},
  title = {{Truncated proposals for scalable and hassle-free simulation-based inference}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2022},
}
Figure