Although many companies collect large volumes of unstructured data, data scientists more frequently analyze and model structured datasets. Data practitioners may continue relying upon structured data until it is easier to compute statistics, which facilitate modeling and decision making, over subsets of large, unstructured dataset. While deep neural networks can be applied to this use case, this approach may be prohibitively computationally expensive. In this paper, Kang et al. use proxy models and stratified sampling to accelerate linear aggregation queries, including with multiple predicates and with group-by keys. Unlike previous work, their method, ABAE, also provides statistical guarantees on query accuracy.