server.logprob() is not normalized

[fritzo]

This is problematic for multiple reasons:

• it corrupts validation statistics based on posterior predictive probabilities
• it is blocking correct implementation of server.correlation() in Implement server.entropy() and server.correlation() #7 .

Some of the test_server_logprob_normalized() tests in server_test.py are marked xfail.

Possible solution

One way to fix this might be to use a different representation of the feature-latent matrices. Right now we're using matrices P(feature|latent), which is likely the problem.

[fritzo]

Fixed by 689d9ea