http://www.cs.toronto.edu/~jmartens/docs/HF_book_chapter.pdf (associated with HFdemo.zip)
https://arxiv.org/pdf/1010.1449v1.pdf (Geodesic acceleration applied to non-linear least squares)
https://arxiv.org/pdf/1503.05671.pdf (updated KFAC method by Martens)
http://www.cs.toronto.edu/~jmartens/research.html (Codes HFDemo.zip and KFAC3-MATLAB.zip)
- Matlab
-
Download data: wget www.cs.toronto.edu/~jmartens/digs3pts_1.mat
-
Open in Matlab: nnet_demo_2.m file
-
Hit Run
- Geodesic acceleration helps reach smaller training error faster than natural gradient on deep networks
- May have to set jacket to 0 or compMode to 'cpu_single' to work.