Subspace projection with PCA and Neural Networks
------------------------------------------------

This project outlines a few methods for subspace project (dimensionality
reduction) typically used in data processing (analytics, informatics, data
science, etc.).

The jumping off point was for a great answer to a question about 
dimensionality reduction on stackoverflow:
https://stats.stackexchange.com/questions/190148/building-an-autoencoder-in-tensorflow-to-surpass-pca
by user 

I've extended this by:
  1) include a CNN architecture, and
  2) using hyperopt for hyperparameter optimisation.

GOAL
----
I would like to produce an improved separation of features in the subspace
(wouldn't we all). This can be visually improved if we add kernel constraints
to the neural networks. How can this be measured and be the objective function?

TODO
----
+ Improve hyperopt integration
+ select between PCA, NN, CNN
+ measure the spread of values in the subspace and compare on that.

RELATED WORK
------------
Reducing the Dimensionality of Data with Neural Networks
G. E. Hinton and R. R. Salakhutdinov, 2006
https://www.cs.toronto.edu/~hinton/science.pdf

NormFace: L_2 Hypersphere Embedding for Face Verification
Feng Wang, Xiang Xiang, Jian Cheng, and Alan L. Yuille, 2017
https://arxiv.org/pdf/1704.06369.pdf