TODO.txt

1. Batch normalization
2. LSTM implementation
3. Refactor the deep belief net code to also use layers: this makes it easy to add fully connected layers to conv nets and to delete the ann implementation, by merging it with the dbn implementation.
4. Consider using fuel for input data
5. Spearmint: do not use the maxEpochs as something that you optimize using spearmint

Debugging wise/theano:
6. Use theano nanguardmode for debugging nans in theano.
7. Use test_value for theano tensor values to ensure that you get shape errors at compile time