2019/09/06 by Christopher Ré, Ré, Christopher, Feng Niu +5 · 3 voices · 3 citations
Computer Science · #Machine Learning and Data Classification #Time Series Analysis and Forecasting #Anomaly Detection Techniques and Applications
paper · pdf · doi:10.48550/arxiv.1909.05372
We describe a system called Overton, whose main design goal is to support\nengineers in building, monitoring, and improving production machine learning\nsystems. Key challenges engineers face are monitoring fine-grained quality,\ndiagnosing errors in sophisticated applications, and handling contradictory or\nincomplete supervision data. Overton automates the life cycle of model\nconstruction, deployment, and monitoring by providing a set of novel\nhigh-level, declarative abstractions. Overton's vision is to shift developers\nto these higher-level tasks instead of lower-level machine learning tasks. In\nfact, using Overton, engineers can build deep-learning-based applications\nwithout writing any code in frameworks like TensorFlow. For over a year,\nOverton has been used in production to support multiple applications in both\nnear-real-time applications and back-of-house processing. In that time,\nOverton-based applications have answered billions of queries in multiple\nlanguages and processed trillions of records reducing errors 1.7-2.9 times\nversus production systems.\n