For Natural image regression with Gaussian features, deeper MLPs achieve similar performance with shorter embeddings than shallower MLPs; shallow models need more features.

来源：paper:§A.2 pp. 12–13; Figure 6 p. 13