I understand how to tie the weights in an autoencoder/decoder, i.e. use the transpose of the encoder weights in the decoder. However, it is unclear what to do with the biases.
If for example,
Encoder:
4 Features -> 3 Hidden -> 2 Hidden
Decoder:
2 Hidden -> 3 Hidden -> 4 Features'
As the biases for a layer are the number of nodes in the next layer they are not symmetric.
In the above example the biases would be 3,2,3,4
However as the Encoder and Decoder are symmetric but in reverse there are 2 biases that are the same size i.e. 3.
Question 1:
Do I tie these similar sized biases?
Question 2:
The ones that I don't tie, if the answer to question 1 is no, or can't tie, the asymmetric ones, are they simply independent variables?