Combating Noisy Labels with Sample Selection by Mining High-Discrepancy Examples