MIT15_097S12_lec15

For example if the documents are all news articles

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: ce). If Θ is a compact set, θ∗ is defined as in (18), A is a neighborhood of θ∗ , and p(θ∗ ∈ A) > 0, then p(θ ∈ A|y ) → 1 as m → ∞. These theorems show that asymptotically, the choice of the prior does not matter as long as it assigns non-zero probability to every θ ∈ Θ. They also show that if the data were generated by a member of the family of likelihood models, we will converge to the correct likelihood model. If not, then we will converge to the model that is ‘closest’ in the KL sense. We give these theorems along with a word of caution. These are asymptotic results that tell us absolutely nothing about the sort of m we encounter in practical applications. For small sample sizes, poor choices of the prior or likelihood model can yield poor results and we must be cautious. 18 Coin Flip Example Part 6. Suppose that the coin flip data truly came from a biased coin with a 3/4 probability of Heads, but we restrict the likelihood model to only include coins with probability in the interval [0, 1/2]: p(y |θ) = θmH (1 − θ)m−mH , θ ∈ [0, 1/2]. This time we will use a uniform prior, p(θ) = 2, θ ∈ [0, 1/2]. The posterior di...
View Full Document

This note was uploaded on 03/24/2014 for the course MIT 15.097 taught by Professor Cynthiarudin during the Spring '12 term at MIT.

Ask a homework question - tutors are online