Convolution methods

Convolution method	Summary	Implementation	Speedup	Reference
FFT	Uses FFT transform to provide faster convolution over filter banks	https://github.com/Theano/Theano/pull/1870	1.5-10x (over convnet) depending on filter and image size	http://arxiv.org/abs/1312.5851
DenseNet	Computes multiscale features from CNN layers	https://github.com/BVLC/caffe/pull/308	10x (over repeated convnet for multiscale)	http://arxiv.org/abs/1404.1869
Low rank	Uses linear decomposition of the filters to reduce work	?	2.5x (no accuracy change) to 4.5x (1% accuracy loss) on their test model	http://arxiv.org/abs/1405.3866v1