ManifoldMix improve the hidden representations and decision boundaries of neural networks at multiple layers by mixing hidden representations rather than input samples.