| 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 |