Interactive demo · Probabilistic Machine Learning: Advanced Topics · Chapter 2

Sensor fusion with unknown noise

Two devices measure the same unknown z, each with an unknown precision. Only the sample mean and the sample variance of each device's readings matter (they are sufficient statistics). The plug-in answer (gold) estimates both precisions by maximum likelihood and reports a single Gaussian; the exact posterior (teal) integrates the precisions out under the priors p(z) ∝ 1 and p(λ) ∝ 1/λ, and is proportional to the product of ((x̄ − z)² + s²)−N/2 over the two devices. When the devices disagree it has two modes: the data cannot say which device is the reliable one.

Device x (red marker)

Device y (green marker)

Try these

"sd" is the sample standard deviation with divisor N, so s² = 0.16 and 0.36 reproduce the readings (1.1, 1.9) and (2.9, 4.1). The plug-in iteration is (2.14)–(2.16): z ← precision-weighted mean, 1/λ ← s² + (mean − z)².

In the lecture notes