Learning minimal abstractions