Learning To Learn via Multi-Stage Optimization

The core of an AI application is a multi-stage program, referred to in industry parlance as a "pipeline". Pipelines are composed of machine learning model code, as well as non-model code used for supporting tasks like data preprocessing, augmentation, post-processing, retrieval, ranking, and serving of inference results. This project investigates multi-stage black-box methods and systems for Learning To Learn in AI pipelines, in order to optimize for multiple simultaneous objectives including task performance, throughput, latency, and cost of compute. The goal is to provide a one-stop system for automatic, hands-free optimization of AI applications according to engineering and business needs.