Learning Time/Memory-Efficient Deep Architectures with Budgeted Super Networks

We propose to focus on the problem of discovering neural network architectures efficient both in terms of prediction quality and cost. For instance, our approach is able to solve the following tasks: 'learn a neural network able to predict well in less than 100 milliseconds' or 'learn an efficient model that fits in a 50 Mb memory'. Our contribution is a novel family of models called Budgeted Super Networks. They are learned using gradient descent techniques applied on a budgeted learning objective function which integrates a maximum authorized cost where this cost can be of different nature. We present a set of experiments on computer vision problems and analyze the ability of our technique to deal with three different costs: the computation cost, the memory consumption cost, and also a \textit{distributed computation} cost. We particularly show that our model can discover neural network architectures that have a better accuracy than the ResNet and CNF architectures on CIFAR-10 and CIFAR-100, at a lower cost.