This preview shows page 1. Sign up to view the full content.
Unformatted text preview: 3 Pick an appropriate choice of priors p (θ|H) given your model. 4 Reason about some actual data and so ﬁnd the posterior over
θ. Nick Jones Inference, Control and Driving of Natural Systems Remark on Priors and Hyperpriors
In the literature there is a great deal of discussion about
appropriate choice of (un/informatative) priors and debate about
the role of conﬁdence intervals etc. . This hasn’t stopped
Bayesian inference being very inﬂuential and widely industrially
applied in Machine Learning.
We are going to hope that correctly characterizing the prior
information that the biological system has been exposed to will
hand us the priors in a natural manner: quantifying the prior
experience of a system is part of describing its structure and
understanding its function. Prior information and model
parameters can have physical interpretations in terms of changes in
RAM or disk magnetization or chemical modiﬁcations.
In practice Bayesians often write their priors p (θ) as parameterized
distributions. These parameters are called hyper-parameters e.g.
the µ and σ in the prior p (θ) ∼ N (σ, µ).
Nick Jones Inference, Control and Driving of Natural Systems Sampling and Posteriors
Here is a crude met...
View Full Document
- Spring '14