Open DifferentSC opened 10 years ago
Thank you for sharing. Actually I feel implementing L-BFGS is quite tricky.
I found this reference and I hope this can help you. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.142.4746
Also there's liblbfgs which is C port of original fortran implementation of lbfgs. http://www.chokkan.org/software/liblbfgs/
It seems that Wikipedia's explanation about L-BFGS algorithm is not enough, so I found a reference from MSRA(Microsoft Research Asia) about implementing it. http://research.microsoft.com/en-us/um/people/yangliu/publication/Fast%20B-spline%20curve%20fitting%20by%20L-BFGS.pdf
Although the article is not about ML, but there is some practical information about how to choose an initial inverse Hessian matrix and appropriate learning rate.
Thanks.