Compressing Features for Learning with Noisy Labels