Confused about the mnist_autoencoder demo

[Werner Van Belle]

Hello,

This is about the mnist_autoencoder example provided in  ./examples/mnist/
- from what I've read an autoencoder uses similar weights between the upgoing and downgoing stages. That is, the weights of decode1 and encode1 should be the inverse of each other. How is this achieved in the demo ?
- training a network like this with multiple layers typically requires a pretraining phase to be performed. Is this automatically done in this example ? If so, how does it happen ?
- why is there a difference between test_on_test and test_on_train ? How does that help the autoencoder ?
- why are there two learningrate parameters in each encoder/decoder ? If these were supposed to helping the pretraining, shouldn't the learningrates differ between each of the stages ?

E.g:
encode1
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }


With kind regards,

Werner,-