In principle yes, but the pre-built binaries require at least compute capability 3. You can build from source, though. When using Ubuntu 16.04 with g++ 5.4 it works out-of-the box using CUDA 8.0.61 and corresponding cuDNN. Things changed with Ubuntu 18.04 and g++ 7, because CUDA 8 and gcc 7 are incompatible. You have to explicitly install an older g++ (e.g. 5.4) and manually enable c++11 features by adding the line

 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")

to CMakeLists.txt. This is required because many libraries in Ubuntu 18.04 are built with gcc 7 with c++11 features enabled by default. Probably you should think about buying a more recent GPU, a GTX1080 is quite cheap, and allows you to use the recent CUDA Toolkit without these hassles!

Short answer: No!

Currently the caffe_unet backend only runs on Linux. caffe_unet is CMake-based, thus you can try to build it using MSVC or mingw for Windows OS. The plugin should work as expected with the Windows build, but without guarantees, it has never been tried! Snapshotting of finetuning processes won't work, because it relies on the Linux kill command for process communication.

Short answer: No!

The virtualization does not allow to fully access all required GPU capabilities. Virtualization must provide information on which parts of the hardware are used by which system. For this virtualbox uses a virtual graphics driver, which cannot be replaced by a "normal" nvidia driver which is not aware of the host system in a virtual environment. For Linux hosts you can theoretically entirely disable the GPU on the host and give dedicated access to the virtual guest system, so that no sharing-issues occur, but I never heard of someone really making this work. On Windows it is simply impossible, because if the Windows OS finds hardware it will use it...