CNN autoencoder

[Alexandre Franco]

Can someone give me an example how to implement an CNN autoencoder?

Thanks.

[Nand Dalal]

I am also working on a similar problem to implement the CAE described here: http://people.idsia.ch/~ciresan/data/icann2011.pdf

I have started by looking at the example in 'examples/mnist/mnist_autoencoder.prototxt', but I am not sure how to "tie" the weights. Namely, I am not sure how to initialize the gradient between the input/middle layer to the gradient between the middle/output layer.

[Youssef Kashef]

Hi Nand and Alexandre,

I was thinking of a solution on a thread with a similar question: How can I run tied weight AutoEncoder ?

The idea is to use the weight sharing feature supported by Caffe and define a new layer class that inherits from InnerProduct that transpose the weights before calling forward() and backward() of the parent class.

Have you tried this or anything similar to it?

thanks,

Youssef