down votefavorite 
<http://stats.stackexchange.com/questions/223573/implementing-contractive-autoencoder-weight-decay#>

i'm trying to implement the idea behind the contractive auto encoder and 
penalty the jacobian of output of a MLP wrt its input to get make it less 
sensitive to noise.

there is a scalar coefficient befor the norm of jacobian term and i was 
wondering how small should i set it to be?

if i set it to more than .001 it will saturate the cells

and if i set the weight decay less than .9999 it will kill the weights.

to clarify its not an autoencoder but an MLP and because its too big i cant 
just train it with noise

i dont know where i'm going wrong? cause .001 seems too small.

-- 

--- 
You received this message because you are subscribed to the Google Groups 
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to