Trying the reproduce the results from (http://jmlr.org/papers/volume11/vincent10a/vincent10a.pdf)
Stacked Denoising Autoencoders: Learning Useful Representations in a Deep Network with a Local Denoising Criterion
Requirements:
- lasagne
- nolearn
The big issue for me is that I am not able to reproduce the filters the authors obtain:
- They get blob/edge detectors
- I get some random crap
Turns out that sample size is the issue: Going from 50000 to 500000 results in the desired filters.
See also this thread