It is well known that variational autoencoders tend to create blurry reconstructions. I want to use the latent variables as image representations, and after training the autoencoder I would like to do transfer learning and use the output of the bottleneck as an input to a binary classifier. I believe the classification task depends on the fine details (high-frequency components) that are lost in the blurry reconstructions. Can I still expect the latent variables to perform well in the classification task?
$\begingroup$
$\endgroup$
2
-
$\begingroup$ By output of bottleneck you mean the latent features (which are penalized to be normally distributed) or output of some decoding layer? $\endgroup$– Jakub BartczukCommented Jun 25, 2018 at 7:34
-
$\begingroup$ @JakubBartczuk I mean the latent features which are penalized to be normally distributed $\endgroup$– elliotpCommented Jun 25, 2018 at 7:36
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
I don't think anything can be said without trying the experiment. On the one hand, if your conjecture is correct and the high-frequency components are important, then I don't see how an auto-encoder that discards the high-frequency components could succeed. On the other hand, your conjecture could be incorrect, and the model could be adequate.
