Learning from higher-order statistics, efficiently: hypothesis tests, random features, and neural networks