Wu Lin comments

Results 30 comments of


                                            Wu Lin

Multiclass Laplace approximation for GPs

this issue is partially solved at https://github.com/shogun-toolbox/shogun/pull/2484 TODO: add implicit gradient wrt hyper-parameter

word2Gaussian

You may have a look at this paper about exponential family embedding. https://arxiv.org/abs/1608.00778

make model selection on GP module more flexible

@karlnapf will work on this issue once `notebooks` are done

make model selection on GP module more flexible

@karlnapf Currently, `pthread` is used to parallelly compute the gradients.

make model selection on GP module more flexible

about moving all GP parameters to log-domain: - not all parameters are positive (for GaussianARDKernel, where weights are FULL_Matrix) (adding a gradient transformation layer is not easy since we have...

make model selection on GP module more flexible

@karlnapf Let me know if you are agree with my idea so that I will implement this idea.

make model selection on GP module more flexible

@karlnapf the existing model selection (`CGradientModelSelection`) does not do gradient transformation, which means there is a bug if a parameter and its gradient are in different domains.

make model selection on GP module more flexible

@karlnapf Any feedback on the model selection

make model selection on GP module more flexible

GPML does keep parameters in log domain if they are positive and parameters in standard domain if they can be negative or zero. UPDATED: (`unconstrained representation`) if we can break...

make model selection on GP module more flexible

If you think adding a new field in `TParameter` is possible, I may need some help from authors' of the class. Adding a new field in `TParameter` is not trivial...