How do we get/define filters in convolutional neural networks?

Question

How to implement a deep autoencoder (eHow do i obtain filters from convulutional neural network(CNN)? My idea is something like this: Do random images of the input images (28x28) and get random patches (8x8). Then use autoencoders to learn the common features of the patches (features = hidden units; approximately 100, for example). Then apply features filters to the input images and do convolution. Am I correct?

I am confused because sometime the literature state only using like, e.g. 8, filters, but in my case I have 100..g. 2 or 3 layers)? Any ideas or resources?

Answer 1

You can follow the tutorial : http://ufldl.stanford.edu/wiki/index.php/UFLDL_Tutorial

This is like a lecture on both auto-encoders and some simple stuff about CNN (convolution and pooling). When you complete this tutorial you will have both auto-encoder implementation and stacked-auto-encoder in your words deep auto-encoder implementation ready.

This tutorial will have exactly what you ask for:

• 28x28 MNIST images

• getting 8x8 patches and learning filters using auto-encoders

• convolving these images by these 8x8 filters

• pooling them

• using the pooled vectors/images and putting them in a soft-max classifier to learn 10 different classes of MNIST database.