Stacked Denoising Auto-Encoder

In Theano. Using the mnist dataset.

Running

python3 train.py

Is the basic run. You can specify more options like this.

python3 train.py n_nodes=[180,42,10] noise=[.1,.05,.025] learning_rate=.1 lambda1=[.5,.1,.1] n_epochs=50 output_folder=plots

In the end you will get a pickle file with three pairs of (w, b) matrices. And a ton of cool pictures.

Dependencies

• Python 3.0 (Yes, not the old dead version! The brand new one in stead!!!)
• Numpy, Scipy, Nose etc. etc.
• Theano
• Pillow

References

• Bengio, Yoshua, Aaron Courville, and Pierre Vincent. "Representation learning: A review and new perspectives." Pattern Analysis and Machine Intelligence, IEEE Transactions on 35.8 (2013): 1798-1828.