AI Glossary / Probabilistic machine learning / Advanced

Variational filtering sequential Monte Carlo

Definition

Variational filtering sequential Monte Carlo refers here to combining particle-based sequential inference with variational optimization of proposals or related model parameters. A particle filter supplies weighted samples, while a variational objective guides learning of the sampling procedure. The sourcebook's variational filtering SMC section groups FIVO, variational SMC and auto-encoding SMC; this entry explains their shared filtering-objective principle.

What it means

A particle filter can represent complicated state uncertainty, but its usefulness depends strongly on where it proposes particles. If most proposed states are inconsistent with the observations, only a few receive useful weight. Variational learning can adjust the proposal to spend computation more effectively.

The proposal is not generally the final posterior approximation. Importance weighting and resampling change the distribution represented by the algorithm. The information available also matters: filtering uses observations available up to the current time, while full-sequence inference or offline proposal training can use a broader history. Calling an algorithm sequential does not automatically mean all its learning occurs online.

How it works

Specify a state-space model, a particle population and a parameterized proposal. At each step, select ancestors according to the chosen resampling scheme, propose new states and compute importance weights from the model-to-proposal density ratio. Keep the support conditions needed for valid importance sampling.

A properly constructed particle filter can provide a nonnegative unbiased estimate Z-hat of the sequence's normalizing constant Z. Jensen's inequality then gives E[log Z-hat] no greater than log Z. This expected log estimate can serve as a tractable variational objective under the method's assumptions.

Optimize proposal or model parameters using an appropriate gradient estimator. Continuous reparameterized draws and discrete ancestor choices require different treatment; simply ignoring a resampling derivative can change the estimator and its guarantees.

The VSMC posterior family includes the particle-generation and selection mechanism, not just a single proposal density. More particles can improve approximation at additional computational cost, but finite runs remain approximate. Streaming adaptation, full-sequence training and alternative particle objectives have distinct protocols and should be identified explicitly.

Three practical examples

1. A nonlinear tracking researcher learns proposals that incorporate current observations, reducing wasted particles in low-likelihood regions.

2. A time-series modeling team trains a latent-state model with a particle-based lower-bound objective and evaluates predictive behavior on held-out sequences.

3. A streaming-inference study adapts proposal parameters as observations arrive, carefully separating that online protocol from offline training using completed sequences.

Think of it this way

Think of sending search teams into plausible areas and then weighting their findings according to how they were sent. Learning improves the dispatch plan, but the final conclusion still depends on the weighting procedure. A convenient plan that never visits a plausible area cannot recover evidence from that area merely by reweighting what it has seen.

Worked example

Consider a one-step, two-state teaching problem with unnormalized target masses 0.1 for A and 0.9 for B. Their normalizing constant is Z = 1. Use proposal probabilities one half for each state and one particle.

The importance weight is 0.1/0.5 = 0.2 when A is sampled and 0.9/0.5 = 1.8 when B is sampled. Its expectation is one, so the normalizer estimate is unbiased. However, its expected logarithm is one half log(0.2) plus one half log(1.8), approximately minus 0.5108. This is below log Z = 0.

Now choose proposal probabilities 0.1 and 0.9. Every importance weight is one, so the expected log estimate is zero and the bound is tight in this toy problem.

With the original equal proposal and two independent particles, a draw containing one A and one B gives average weight one and log estimate zero. That one realization does not prove the expected bound is tight. Considering all four ordered draws gives expected log estimate one quarter log(0.2) plus one quarter log(1.8), approximately minus 0.2554.

This is a first-step importance-sampling special case, with no ancestor resampling yet. It isolates the objective's logic rather than reproducing a complete sequential algorithm or its gradient optimization.

Illustrative case study

Fictional case: Alder Monitoring models a changing hidden equipment state from nonlinear measurements. A basic transition proposal places most particles where the latest observation is unlikely, so analysts investigate a learned observation-aware proposal.

They compare effective sample sizes, state-estimation errors and processing time on recorded sequences. They also test a streaming protocol separately from offline training. An improved training bound is useful evidence about the optimization objective, but the deployment decision depends on reliable filtering under the actual information and computation constraints.

Common misconceptions and mistakes

An unbiased normalizer estimate does not imply an unbiased log-normalizer estimate. The logarithm is nonlinear, as the worked example shows.

A lower bound holds for the expectation under the specified sampling procedure; an individual random log estimate need not lie below the true log normalizer. A learned proposal also does not eliminate importance weighting or automatically prevent particle collapse.

Filtering, smoothing and online parameter learning are different tasks. Avoid describing all methods with particle-based variational objectives as the same algorithm. The sourcebook uses this heading for a family of methods, rather than one uniquely specified algorithm.

Business use

Business value is potential improvement in sequential probabilistic inference when a fixed particle budget is important. Teams should assess the cost of learning the proposal as well as the cost of running it, and evaluate uncertainty, latency and robustness to unfamiliar sequences. Preserve a clear record of the objective, resampling rule, proposal inputs and training protocol so results can be interpreted and reproduced.

People also ask

Is the proposal itself the posterior?

Generally no; weighting and selection are part of the approximation.

Must learning happen online?

No. Some approaches train on complete sequences and later run sequentially.

Does a higher observed sample objective prove improvement?

A noisy realization alone does not establish an improvement in its expectation or in task performance.

Check your understanding

Why is the one-particle log objective negative despite an unbiased estimate of one?

Answer: Jensen's inequality separates the expectation of a logarithm from the logarithm of an expectation.

What makes the better proposal exact in the example?

Answer: It equals the normalized target, making every importance weight constant.

Related terms

• Importance sampling
• Proposal distribution
• variational sequential Monte Carlo
• particle filter
• filtering variational objective
• resampling and streaming variational Monte Carlo

Learn next

Next, extend the example to two time steps and track ancestor indices and incremental weights. Then examine gradient estimators for discrete resampling and distinguish online filtering objectives from full-sequence training objectives.

More in Probabilistic machine learning

• Unscented transform - The unscented transform is a deterministic method for approximating the mean and covariance of a random quantity after a nonlinear transformation.
• Variational Bayes expectation maximization - Variational Bayes expectation maximization, often called variational Bayesian EM or VBEM, is an iterative approach to approximate Bayesian inference that...
• Variational expectation maximization - Variational expectation maximization, or variational EM, is an alternating optimization method for latent-variable models when exact posterior inference is...
• Variational message passing - Variational message passing, or VMP, is a way to organize variational inference as local computations and messages in a probabilistic graphical model.
• Variational recurrent neural network - A variational recurrent neural network, or VRNN, is a sequence model that combines recurrent neural-network memory with latent random variables at each time...
• Vector-Jacobian product - A vector-Jacobian product, or VJP, combines sensitivities at a function's outputs with its derivatives to produce sensitivities at its inputs.

Sources

Primary sources: Murphy, Probabilistic Machine Learning: Advanced Topics, Section 13.6.7, pages 569–570: https://github.com/probml/pml2-book/releases/latest/download/book2.pdf ; Naesseth, Linderman, Ranganath and Blei, Variational Sequential Monte Carlo: https://proceedings.mlr.press/v84/naesseth18a/naesseth18a.pdf . Full sourcebook downloaded and passage read independently, resolving earlier access and title uncertainty. The section groups FIVO, VSMC and auto-encoding SMC. Examples are original.

Previous: Variational expectation maximization Next: Variational lower bound