Feature extraction using convolution

I'm following this link to implement convolution-based feature extraction,

Link: http://ufldl.stanford.edu/wiki/index.php/Exercise:Convolution_and_Pooling

Questions

1) Why is the autoencoder applied on 8x8 image samples to get the optimal weights, by which filters of 8x8 are extracted and convoluted on 64 x 64 image samples?

2) On what criteria the 8x8 image samples are selected for the autoencoder? Let's say the only images given are of size 64x64, then how would the filters be created? If we apply autoencoders on 64x64 images, we will get filters of the same size which in convolution would be implausible.

Thanks in advance...