Sampling-based Algorithms for Fast and Deployable AI

Abstract

We present sampling-based algorithms with provable guarantees to alleviate the increasingly prohibitive costs of training and deploying modern AI systems. At the core of this thesis lies importance sampling, which we use to construct representative subsets of inputs and compress machine learning models to enable fast and deployable systems. We provide theoretical guarantees on the representativeness of the generated subsamples for a variety of objectives, ranging from eliminating data redundancy for efficient training of ML models to compressing large neural networks for real-time inference. In contrast to prior work that has predominantly focused on heuristics, the algorithms presented in this thesis can be widely applied to varying scenarios to obtain provably competitive results. We conduct empirical evaluations on real-world scenarios and data sets that demonstrate the practicality and effectiveness of the presented work.