next up previous contents
Next: Noise models Up: BAYESIAN LEARNING IN PRACTICE Previous: Bits-back argument

Specification of the model and priors

So far we have discussed the rules which the learning should ideally follow and some of the practical approximation for these rules. This is roughly the point where the theory ends and practice begins. Bayesian probability theory tells how the beliefs in propositions should be adapted, but it does not indicate what exactly the propositions should be. Neither does it specify the prior beliefs of the learning system. These choices are left for the designer and depend on the problem at hand. This section describes some of the rules of thumb which have been found to be useful guidelines for the design of a learning system (see, e.g., [75,103,28]).


Harri Valpola