We need to build a proper python package-structure for the project with init.py files. Name resolving works inside a package.

Currently code fails if THEANO_FLAGS='floatX=float32' is not set.

Either support both or document it well.

Currently the test_runner.py fails and is no longer maintained in theano branch.

For linear system, optimal dimensionality calculation gave 1 for MNIST dataset. Applying an 1 dimensional system to this would pull the dataset closer to origin. Doing the optimal dim calculation again for the concatenated (ubar, xbar) input, we would find a new, higher n for the new system. Repeating this could potentially be used for automatic data preprocessing.

Moreover, after implementing dynamic dimensionality reconfiguration (#1), it could be done even faster.

Theory says it should, so try to make loops numerically more stable.

When increasing and decreasing dimensions, it's good to interpolate X_i from and to 0, so that it wouldn't produce too turbulent changes to the system.