Coming from #166.

1. The various random number generators are not thread safe. It might make sense to make them such. Are there cases against not making them such?
2. Should the generators be seeded with cryptographic random seed to make sure they're random when no seed is given?
3. The MachineLearning.Net has tendency to put many classes in one file (and have them uncommented). Would there be opposition to put the classes into separate files?
4. The library has also a tendency to put functions not strictly needed for the functionality in the domain classes. Such as saving, like at

   machinelearning/src/Microsoft.ML.Core/Utilities/Random.cs

   Line 186 in 436700a

   public void Save(BinaryWriter writer)

   . Would there be opposition to provide these as extension methods or in some other fashion? Since it might be one might want to serialize things differently and organizing differently would avoid baking in dependencies. A related idea: IFormatProvider in TextLoader #126 (comment).

At a glance point, there can be other issues too. Would there be opposition to refactor points 1 through 3? Possibly 4?

The source: https://github.com/dotnet/machinelearning/blob/436700aadf615e7f05a22925476cc441c63a919d/src/Microsoft.ML.Core/Utilities/Random.cs