Uncertainty-aware self-training with expectation maximization basis transformation