Towards Reliable Simulation-Based Inference with Balanced Neural Ratio Estimation
Towards Reliable Simulation-Based Inference with Balanced Neural Ratio Estimation
复制标题
通过平衡神经比率估计实现可靠的基于模拟的推理
DOI：
10.48550/arxiv.2208.13624
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Gilles Louppe
中科院分区：
文献类型：
--
作者：
Arnaud Delaunoy;Joeri Hermans;François Rozet;Antoine Wehenkel;Gilles Louppe
基于模拟的推理的现代方法依赖于深度学习代理来实现计算机模拟器的近似推理。然而，在实践中，估计后验的计算忠实度很少得到保证。例如，赫尔曼斯等人。 （2021）表明，当前基于模拟的推理算法可能会产生过度自信的后验，因此存在错误推理的风险。在这项工作中，我们引入了平衡神经比率估计（BNRE），这是 NRE 算法的一种变体，旨在产生更保守的后验近似值，从而提高其可靠性，同时共享相同的贝叶斯最优解。我们通过强制平衡条件来实现这一目标，该平衡条件增加了小型模拟预算制度中的量化不确定性，同时随着预算的增加仍收敛于精确的后验。我们提供的理论论据表明 BNRE 倾向于产生比 NRE 更保守的后验代理。我们在各种任务上评估 BNRE，并表明它在所有测试基准和模拟预算上产生保守的后验代理。最后，我们强调 BNRE 比 NRE 更容易实现，并且不会引入任何计算开销。
Modern approaches for simulation-based inference rely upon deep learning surrogates to enable approximate inference with computer simulators. In practice, the estimated posteriors' computational faithfulness is, however, rarely guaranteed. For example, Hermans et al. (2021) show that current simulation-based inference algorithms can produce posteriors that are overconfident, hence risking false inferences. In this work, we introduce Balanced Neural Ratio Estimation (BNRE), a variation of the NRE algorithm designed to produce posterior approximations that tend to be more conservative, hence improving their reliability, while sharing the same Bayes optimal solution. We achieve this by enforcing a balancing condition that increases the quantified uncertainty in small simulation budget regimes while still converging to the exact posterior as the budget increases. We provide theoretical arguments showing that BNRE tends to produce posterior surrogates that are more conservative than NRE's. We evaluate BNRE on a wide variety of tasks and show that it produces conservative posterior surrogates on all tested benchmarks and simulation budgets. Finally, we emphasize that BNRE is straightforward to implement over NRE and does not introduce any computational overhead.
影响因子：
0.700
作者：
Yu. N. Vershinin;D. S. Il’ichev;P. A. Morozov
通讯作者：
P. A. Morozov