#bias-term

Question 1: Select all that apply

Documents are binary word-indicator vectors $\Phi$ , and a logistic regression model $P(y = 1 \mid \Phi, w) = g(w^\top\Phi)$ without a bias term is trained on $n$ labelled documents. You may then request labels for the other documents one at a time, using the uncertainty $2p_t(1 - p_t)$ with $p_t = P(y = 1 \mid \Phi_t, \hat w)$ . Which statements are true? Select all that apply.

A word that occurs in every document effectively acts as a bias term1
A word that occurs only in the $n$ initially labelled documents works just as well as a bias term2
After each query, the uncertainty of every unlabelled document decreases or stays the same3
The next query should be the document with the largest $2p_t(1 - p_t)$ 4

Timed test