6.891 HW #3 Clarifications
-
-
- "quarantees" should be "guarantees"
-
-
-
-
-
-
- The update for \alpha_k is stated incorrectly. It should be \alpha_k = \min (1,1/2 \log[(1-\epsilon_k)/\epsilon_k]). However, we won't be taking off points if you use the \alpha_k update specified in 1-II-c.
- Discuss how the results were affected by the change in \alpha_i (if they were affected) and discuss (in general) how such a change might affect test performance.
-
-
LOOCV is unbiased (the statement of the question is not a mistake). Give us an argument for why this is true.
-
-
- 0.001/max(...) should be 0.001*max(...)
-
-
-
-
- The \theta_0 subscript is there to indicate that \theta_0 is
held constant for the expectation. It is not there to indicate
that the expectation is over \theta_0. In (more or less) plain
English, Eq. 14 should be read "sum over i, expected value of \log
P(x_i, y_i|\theta) over the distribution P(y_i|x_i, \theta_0)".
-
-
-
-
- The initially released versions of inimix.m, runmix.m,
find_stump.m and boost.m don't work with Matlab v5.2. Updated
versions have been posted.
-
- Create two models: One using one component per class (same
thing as a mixture of two Gaussians) and one using two components
per class where EM is run separately for each class to determine
the parameters for the two Gaussians within that class.
-