All the weights, and many training sets are available under http://www.cs.nyu.edu/~zaremba/data .

cd .git/hooks
.git/hooks$ ln -s ../../theano/PRESUBMIT.py pre-commit
chmod a+x ./pre-commit

Wojciech:

• Write loss (replace cost), write L2 loss, and cross-entropy. Place a L2 penalty on weights.

Emily:

• Implement normalizations.
• Import CIFAR-10.
• Get baseline on cifar-10
• get a simple regularization through approx (schedule it with saving, and testing)

Joan:

• Get nuclear norm to work

Potential:

• More parameters for starting execution (there is special lib in python to parse it arg-something).
• Save mnist in float8.
• Mean subtraction layer, division layer, subtraction layer. Combine them to input layer.
• Add data_augmentation layers (like random view).
• Write script to visulalize weights from the first layer.
• Common initialization mechanism.
• Import Pierre's model.
• add some tests (also end2end test) !! Move end2end from other project.
• model should save its model configuration as it runs