- We do tansformation on those fartherest points which should be the supporter of BVM.
- We add the new point in the end of prob->x
- We allcote the memory in advance by using the param.times optionion. So the number of support vectors couldn’t be larger than param->times*prob->l.
- when a new point is added in, we need to update the kernel matrix as well, mainly for the diagonal element calculation.
- For the storage, we can allocate more depending on our use. But we should notice that the memory can not be overlapped.
The framework for invariant learning using BVM: