Amortized Monte Carlo Integration
Introduction
In Bayesian machine learning, the ultimate goal is often not just to recover a posterior distribution $p(x|y)$, but to compute an expectation over that posterior: $\mathbb{E}_{p(x|y)}[f(x)]$. This is critical for tasks such as making predictions, calculating risk, or performing decision-making under uncertainty. Amortized Monte Carlo Integration (AMCI) is a novel framework designed to directly amortize these calculations. Unlike traditional methods that approximate the posterior in a target-agnostic way, AMCI is "target-aware," tailoring its internal proposals to the specific function $f(x)$ being integrated.
The Problem with Target-Agnostic Inference
Current amortized inference techniques, such as Variational Autoencoders (VAEs), typically learn a single proposal distribution to approximate the posterior. This approach is target-agnostic; the same approximation is used regardless of the downstream function $f(x)$. This is inherently inefficient if $f(x)$ is known beforehand.
Furthermore, these methods generally rely on Self-Normalized Importance Sampling (SNIS). SNIS has a fundamental theoretical limitation: a "hard floor" on its accuracy. Even with a perfect proposal that matches the posterior, the estimator's variance remains non-zero for any non-constant function. This bound limits the precision achievable with a fixed number of samples.
Amortized Monte Carlo Integration (AMCI)
AMCI shifts the focus from posterior approximation to expectation estimation. It breaks the expectation ratio into three distinct components and learns a dedicated amortized proposal for each:
- $q_1^+$: Tailored to the positive part of the unnormalized expectation ($f^+(x)p(x,y)$).
- $q_1^-$: Tailored to the negative part ($f^-(x)p(x,y)$).
- $q_2$: Tailored to the normalization constant (the evidence, $p(y)$).
By estimating the numerator and denominator separately with these optimal proposals, AMCI can theoretically achieve zero-variance estimates with as little as one sample from each component. This "component-wise amortization" allows AMCI to bypass the SNIS bound.
Experimental Evaluation
The authors evaluated AMCI on several challenging tasks:
- Tail Integral Calculation: Estimating probability mass in the tails of 1D and 5D Gaussian distributions. These regions have very low prior probability, making them difficult for standard samplers.
- Cancer Treatment Planning: A simulator-based task using an ODE model of tumor growth. The goal is to predict expected treatment loss based on noisy tumor size observations.
Key Results and Insights
The most significant result is AMCI's ability to "break" the theoretical SNIS bound. As shown in the tail integral experiments, AMCI achieves levels of accuracy that are mathematically impossible for standard SNIS-based amortized methods.
In the cancer treatment planning task, AMCI demonstrated extreme sample efficiency, reaching the same accuracy as standard amortized baselines while using 10,000x fewer samples. This efficiency stems from its target-awareness: if the function $f(x)$ is large in areas where the posterior is small (e.g., in the tails), AMCI's proposals will specifically target those regions, whereas a posterior-matching proposal would miss them entirely.
Conclusion
AMCI represents a paradigm shift in amortized inference by moving from target-agnostic posterior matching to target-aware expectation estimation. By breaking the SNIS bound and providing a framework for amortizing over families of functions, it offers a powerful tool for high-precision Bayesian computation in complex models. While it introduces additional training overhead by requiring three networks, the dramatic gains in online sample efficiency and accuracy make it a compelling choice for many real-world applications.
Summary updated