26.10.2013 Views

Nonparametric Bayesian Discrete Latent Variable Models for ...

Nonparametric Bayesian Discrete Latent Variable Models for ...

Nonparametric Bayesian Discrete Latent Variable Models for ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

4.5 A Choice Model with Infinitely Many <strong>Latent</strong> Features<br />

Figure 4.11: Examples of reconstructed images and their latent features. Last column: original<br />

digits, second last column: reconstructed digits, other columns: features used <strong>for</strong><br />

reconstruction. For each image, the binary vector zi determines the presence or<br />

absence of a feature. The real valued vector zi determines how much each feature<br />

contributes to the observed image.<br />

log likelihood<br />

m k (# objects with feature k)<br />

×10 5<br />

−4<br />

−4.1<br />

−4.2<br />

−4.3<br />

300<br />

200<br />

100<br />

0<br />

2500 5000<br />

iterations<br />

7500 10000<br />

20 40 60 80 100 120<br />

k (feature label)<br />

# iterations<br />

# objects<br />

400<br />

300<br />

200<br />

100<br />

0<br />

50 100<br />

# active feats<br />

150<br />

150<br />

100<br />

50<br />

0<br />

5 10<br />

# active feats<br />

15<br />

Figure 4.12: Distribution of the number of active features over 10000 iterations (left). Distribution<br />

of number of active features <strong>for</strong> each input at one iteration (right). Note<br />

that the number of features that a particular data point has is much less than<br />

the total number of active features. The number of observations possessing each<br />

feature at one iteration. Note that although the total number of active features is<br />

much larger, about half of the features belong to only a few observations.<br />

101

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!