Adaptive Sampling for Deep Learning via Efficient Nonparametric Proxies