1 Section 1. Course introduction and prerequisites
2021-08-17
1.1 Resources
- BDA chapter 1 and reading instructions
- lectures:
- video: ‘Introduction to the course contents’
- slides
- Assignment 1
1.2 Notes
1.2.1 Reading instructions
- model vs. likelihood for \(p(y|\theta, M)\)
- model when the function is in terms of \(y\)
- should be written as \(p_y(y|\theta, M)\)
- used to describe uncertainty about \(y\) given values of \(\theta\) and \(M\)
- likelihood when the function is in terms of \(\theta\)
- should be written as \(p_\theta(y|\theta, M)\)
- the posterior distribution describes the probability for different values of \(\theta\) given fixed values for \(y\)
- “The likelihood function is unnormalized probability distribution describing uncertainty related to \(\theta\) (and that’s why Bayes rule has the normalization term to get the posterior distribution).”
- model when the function is in terms of \(y\)
- exchangeability
- independence is stronger condition than exchangeability
- independence implies exchangeability
- exchangeability does not imply independence
- exchangeability is related to what information is available instead of the properties of unknown underlying data generating mechanism
1.2.2 Lecture notes
Introduction to uncertainty and modelling
- two types of uncertainty:
- aleatoric: due to randomness
- epistemic: due to lack of knowledge
- model vs. likelihood:
- model
- \(p_y(y|\theta, M)\)
- a function of \(y\) given fixed values of \(\theta\)
- describes aleatoric uncertainty
- likelihood
- \(p_\theta(y|\theta, M)\)
- function of \(\theta\) given fixed values of \(y\)
- provides information about the epistemic uncertainty
- is not a probability distribution
- Bayes rule combines the likelihood with prior uncertainty to update the posterior uncertainty
- example with a bag containing red and yellow chips:
- probability of red = #red / #red + #yellow = \(\theta\)
- \(p(y = \text{red} | \theta)\): aleatoric uncertainty
- predicting the probability of pulling a red chip has uncertainty due to randomness even if we new \(\theta\) exactly
- \(p(\theta)\): epistemic uncertainty
- we don’t know \(\theta\) but could compute it exactly if we knew the contents of the bag
- model
Introduction to the course contents
- benefits of Bayesian approach
- integrate over uncertainties to focus on interesting parts
- use relevant prior information
- hierarchical models
- model checking and evaluation