Naming.
Using of namespaces is strongly recommended, especially in the public project. And it should not be OpenNNL unless you are not going much deeper with developing you implementation and creating the comunity around it.
The class should be named accordingly the entity it represents, eg. NeuralNetwork.
Comments.
You should use english comments in public projects. And i personally recommend doing so in private ones, too.
Add somewhere a comment with reference to description of model you are using or the file name with its description.
You use very little comments in implementation part. This decision is totally up to you, but using comments on what the code is doing is always better then not using them in my opinion.
Interface
Public methods before private methods are preferrable, because they are the ones the programmer is looking for in most cases. Also, consider applying PIMPL idiom.
XXXRefs methods are breaking class invariant. Also see my notes on arrays below.
Plain C arrays in interface are not something your users will be happy to have. Consider using std::vector, overloading operator[] and building more abstractions, eg NeuronLayer. Also see notes on arrays below.
load/saveNetwork simply should not be there. The class is responsible for modelling network, not disk operations. Consider overloading operators >>
I recommend to hide compute methods from user. Instead, make getOutputs to compute the result. Of course, use dirty flag to indicate if it should be comuted at all.
BP and IDBD should be decoded in comments, i think.
The code
Extensive use of plain C arrays. There should be a strong reason for this. Standard classes std::vector and std::valarray is less error prone and easer to use. If used properly thy have no or very little overhead. As i see, you considered to provide CUDA implementation, which may restrict in usage of theese classes, but consider providing vectors in interface anyway.
What about copy and assignment? Be explicit in this field, implement operations or make them inaccesible.
Note about inline functions. It's strange to have no implementation in header. And, in my experience such functions will be inlined by compiler without inline modifier. Functions compiler won't inline, it will not inline with modifier. In my opinion, this keyword is obsolette.
calculateWorker is not very expressive name. Consider change it and possibly devide this function on several functions doing descibable tasks.
Consider refactoring changeWeightsXXX functions. They are too long, not commented and have commented debug code in them.
Functions doEpochXXX have debug code in them.