We present RecD (Recommendation Deduplication), a suite of end-to-end infrastructure optimizations across the Deep Learning Recommendation Model (DLRM) training pipeline. RecD addresses immense storage, preprocessing, and training overheads caused by feature duplication inherent in industry-scale DLRM training datasets. Feature duplication arises because DLRM datasets are generated from interactions. While each user session can generate multiple training samples, many features' values do not change across these samples. We demonstrate how RecD exploits this property, end-to-end, across a deployed training pipeline. RecD optimizes data generation pipelines to decrease dataset storage and preprocessing resource demands and to maximize duplication within a training batch. RecD introduces a new tensor format, InverseKeyedJaggedTensors (IKJTs), to deduplicate feature values in each batch. We show how DLRM model architectures can leverage IKJTs to drastically increase training throughput. RecD improves the training and preprocessing throughput and storage efficiency by up to 2.48x, 1.79x, and 3.71x, respectively, in an industry-scale DLRM training system.

In this dissertation, I derive a new method to estimate the Vapnik-Chervonenkis Dimension (VCD) for the class of linear functions. This method is inspired by the technique developed by Vapnik et al. Vapnik et al. (1994). My contribution rests on the approximation of the expected maximum difference between two empirical Losses (EMDBTEL). In fact, I use a cross-validated form of the error to compute the EMDBTEL, and I make the bound on the EMDBTEL tighter by minimizing a constant in of its right upper bound. I also derive two bounds for the true unknown risk using the additive (ERM1) and the multiplicative (ERM2) Chernoff bounds. These bounds depend on the estimated VCD and the empirical risk. These bounds can be used to perform model selection and to declare with high probability, the chosen model will perform better without making strong assumptions about the data generating process (DG). I measure the accuracy of my technique on simulated datasets and also on three real datasets. The model selection provided by VCD was always as good as if not better than the other methods under reasonable conditions.

We introduce the speculate-correct method to derive error bounds for local classifiers. Using it, we show that k-nearest neighbor classifiers, in spite of their famously (fractured decision boundaries, have exponential error bounds (k) with O lnn)/n error bound range for n in-sample examples. Keywords: nearest neighbors, statistical learning, supervised learning, error bounds, generalization 2000 MSC: 62G99, 2000 MSC: 68Q32, 2000 MSC: 62M99 1. Introduction Local classifiers use only a small subset of their examples to classify each input. The best-known local classifier is the nearest neighbor classifier. To classify an example, a k-nearest neighbor (k-nn) classifier uses a majority vote over the k in-sample examples closest to the example. We assume k is odd, and we assume binary classification.