Title:Building effective deep neural network architectures one feature at a time

Abstract:Successful training of convolutional neural networks is often associated with the training of sufficiently deep architectures composed of high amounts of features while relying on a variety of regularization and pruning techniques to converge to less redundant states. We introduce an easy to compute metric, based on feature time evolution, to evaluate feature importance during training and demonstrate its potency in determining a networks effective capacity. In consequence we propose a novel algorithm to evolve fixed-depth architectures starting from just a single feature per layer to attain effective representational capacities needed for a specific task by greedily adding feature by feature. We revisit popular CNN architectures and demonstrate how evolved architectures not only converge to similar topologies that benefit from less parameters or improved accuracy, but furthermore exhibit systematic correspondence in representational complexity with the specified task. In contrast to conventional design patterns that typically have a monotonic increase in the amount of features with increased depth, we observe that CNNs perform better when there is a peak in learnable parameters in intermediate, with falloffs to earlier and later layers.