Generalizing MLPs With Dropouts, Batch Normalization, and Skip Connections

A multilayer perceptron (MLP) is typically made of multiple fully connectedlayers with nonlinear activation functions. There have been several approachesto make them better (e.g. faster convergence, better convergence limit, etc.).But the researches lack structured ways to test them. We test different MLParchitectures by carrying out the experiments on the age and gender datasets.We empirically show that by whitening inputs before every linear layer andadding skip connections, our proposed MLP architecture can result in betterperformance. Since the whitening process includes dropouts, it can also be usedto approximate Bayesian inference. We have open sourced our code, and releasedmodels and docker images at https://github.com/tae898/age-gender/