Attempt at an implementation of http://arxiv.org/pdf/1311.1780v7.pdf in Theano/Lasagne.
Make sure you have Lasgane and Theano properly installed. Then:
python mnist_test.py
Scan function/symbolic tensor functions are not optimized! Takes way too long on a CPU. Next step, try one of the following:
- Refactor code to work in GPU, test runtime
- Consider alternative implementation that overrides the Pool and PoolGrad op classes.