Neural posterior estimation (NPE) is a popular Bayesian simulation-based inference (SBI) approach that uses neural networks trained on simulated data--parameter pairs to approximate the posterior given a user-chosen prior. For high-dimensional observations, NPE methods typically approximate the conditional density of the parameter given a user-chosen or learned summary statistic, rather than the entire dataset, due to the curse of dimensionality. However, both approaches face challenges: hand-crafting good summaries for complex models requires considerable effort from scientists, while learning them from massive datasets can be prohibitively computationally expensive. To address these challenges, we propose approximating the conditional density of a parameter given low-dimensional sketches of the data. We show that, for a model with a p-dimensional parameter, only 2p+1 random sketches are needed for posterior consistency. We compare with state-of-the-art embedding networks for NPE, and we find that our method matches the best embedding network's performance while increasing computational efficiency.
Figures & tables
Scalable
Automated
Interpretable
Hand-crafted
✓
×
✓
Learned
∼
✓
×
Randomized
✓
✓
×
Table 1: Comparison of paradigms for summary statistics. The table displays yes ( ✓ ), depends on method ( ∼ ), and no ( × ).
Figure 1: Moving average with g-and-k innovations. Marginal posterior densities from each method at different sample sizes.
Figure 2: Autoregressive with g-and-k innovations. Marginal posterior densities from each method at different sample sizes.
Figure 3: State-space with 25-dimensional observations. Marginal posterior densities from each method at different sample sizes.
Figure 4: Stochastic volatility with regime-switching. Marginal posterior densities from each method at different sample sizes.
Figure 5: Comparison of posterior density quality . Wasserstein-1 distance to δθ0 at different sample sizes, averaged over 1000 trajectories simulated under Pθ0,n .
Figure 6: Comparison of total computation cost. Preprocessing and training time at different sample sizes.
Figure 7: Comparison of optimization speed. Adam optimizer update time at different sample sizes.