ExpandNets: Linear Over-parameterization to Train Compact Convolutional Networks

We introduce an approach to training a given compact network. To this end, we leverage over-parameterization, which typically improves both neural network optimization and generalization.