uncertainty/deep-ppl.md

@@ -26,6 +26,16 @@ The central problem of PPLs is that comptuting the posterior distribution of lat
 
 Use variational Inference
 
+** How to Use Variational Inference?**
+
+The goal is to find q<sup>*</sup>(θ) from a family of simpler distributions Q over the latent variable θ that is closest to the true posterior p(θ|x<sub>1</sub>, ..., x<sub>n</sub>).
+This is done by:
+1.  Define model (e.g. a multi-layer perceptron; MLP) and initialise priors
+2.  Specify guides (simple probabilistic distributions over the latent variable θ
+3.  Run inference, i.e. learn distributions that fit the observed data -> posterior distribution for θ
+4.  Sample concrete weights and biases from θ (e.g. results in a concrete instance of the MLP)
+5.  (Optional) repeat weights and biases sampling to generate multiple concrete model instances
+6.  (Optional) use voting to get the best model (e.g. MLP) from the multiple samples
 
 **Why Deep Learning and PPL?**