VBF Experiments, April 2026 on mlbot.blog

A Tour Of Learned And Reference-Free Bayesian Filters

Thu, 30 Apr 2026 14:03:11 +0530

This is a long technical note about a small research program in Bayesian filtering. The starting point is familiar if you know Kalman filters: there is a hidden state, there are noisy measurements, and the filter has to update its belief online.

Reference-Free Quadrature Filters For The Sine Benchmark

Thu, 30 Apr 2026 13:18:00 +0530

The last part of the week stepped away from amortized training and asked a sharper question:

How far can a deterministic, reference-free filtering update go if it directly integrates the known nonlinear likelihood and projects the result back to a strict family?

When Mixtures Beat Local ELBO In Nonlinear Filtering

Thu, 30 Apr 2026 13:17:00 +0530

The nonlinear objective-repair branch left a clear gap: the best fully unsupervised strict Gaussian filter improved robustness, but still had weak coverage and very low variance ratios. The next branch tested whether that was an objective problem, a posterior-family problem, or a coupled problem.

Repairing a Nonlinear Strict Filter Without Reference Targets

Thu, 30 Apr 2026 13:16:00 +0530

After the scalar benchmark, the work moved to a nonlinear sine-observation model:

\[ z_t = z_{t-1} + w_t,\quad w_t \sim \mathcal{N}(0,Q) \]\[ y_t = x_t \sin(z_t) + v_t,\quad v_t \sim \mathcal{N}(0,R) \]

The strict filtering contract stayed the same:

\[ q^F_t = \operatorname{update}(q^F_{t-1}, x_t, y_t) \]

No hidden sequence state was allowed in the headline rows. The filter had to export an explicit online filtering marginal at each time step.

Variational Filtering, Rebuilt From the Linear Case

Thu, 30 Apr 2026 13:15:00 +0530

This week started by rebuilding the variational Bayesian filtering experiments around a scalar linear-Gaussian state-space model. The goal was not to win on a toy problem. The goal was to make the mechanics testable before asking nonlinear questions:

\[ z_t = z_{t-1} + w_t,\quad w_t \sim \mathcal{N}(0, Q) \]\[ y_t = x_t z_t + v_t,\quad v_t \sim \mathcal{N}(0, R) \]

The filter carried a strict online marginal \(q^F_t(z_t)\) plus an edge/backward conditional \(q^B_t(z_{t-1} \mid z_t)\). That made the posterior edge factor explicit: