Brian Hentschel

Brian Hentschel, Peter J. Haas, Yuanyuan Tian

To maintain the accuracy of supervised learning models in the presence of evolving data streams, we provide temporally-biased sampling schemes that weight recent data most heavily, with inclusion probabilities for a given data item decaying over time according to a specified "decay function". We then periodically retrain the models on the current sample. This approach speeds up the training process relative to training on all of the data. Moreover, time-biasing lets the models adapt to recent changes in the data while---unlike in a sliding-window approach---still keeping some old data to ensure robustness in the face of temporary fluctuations and periodicities in the data values. In addition, the sampling-based approach allows existing analytic algorithms for static data to be applied to dynamic streaming data essentially without change. We provide and analyze both a simple sampling scheme (T-TBS) that probabilistically maintains a target sample size and a novel reservoir-based scheme (R-TBS) that is the first to provide both control over the decay rate and a guaranteed upper bound on the sample size. If the decay function is exponential, then control over the decay rate is complete, and R-TBS maximizes both expected sample size and sample-size stability. For general decay functions, the actual item inclusion probabilities can be made arbitrarily close to the nominal probabilities, and we provide a scheme that allows a trade-off between sample footprint and sample-size stability. The R-TBS and T-TBS schemes are of independent interest, extending the known set of unequal-probability sampling schemes. We discuss distributed implementation strategies; experiments in Spark illuminate the performance and scalability of the algorithms, and show that our approach can increase machine learning robustness in the face of evolving data.

Abdul Wasay, Brian Hentschel, Yuze Liao et al.

Ensembles of deep neural networks significantly improve generalization accuracy. However, training neural network ensembles requires a large amount of computational resources and time. State-of-the-art approaches either train all networks from scratch leading to prohibitive training cost that allows only very small ensemble sizes in practice, or generate ensembles by training a monolithic architecture, which results in lower model diversity and decreased prediction accuracy. We propose MotherNets to enable higher accuracy and practical training cost for large and diverse neural network ensembles: A MotherNet captures the structural similarity across some or all members of a deep neural network ensemble which allows us to share data movement and computation costs across these networks. We first train a single or a small set of MotherNets and, subsequently, we generate the target ensemble networks by transferring the function from the trained MotherNet(s). Then, we continue to train these ensemble networks, which now converge drastically faster compared to training from scratch. MotherNets handle ensembles with diverse architectures by clustering ensemble networks of similar architecture and training a separate MotherNet for every cluster. MotherNets also use clustering to control the accuracy vs. training cost tradeoff. We show that compared to state-of-the-art approaches such as Snapshot Ensembles, Knowledge Distillation, and TreeNets, MotherNets provide a new Pareto frontier for the accuracy-training cost tradeoff. Crucially, training cost and accuracy improvements continue to scale as we increase the ensemble size (2 to 3 percent reduced absolute test error rate and up to 35 percent faster training compared to Snapshot Ensembles). We verify these benefits over numerous neural network architectures and large data sets.