Andrew Haigh

Title: N_batch=1: Going from proof-of-concept to realtime inference

Abstract:

Publication provides a certain set of priorities and motivation when it comes to the research and development of machine learning models which may differ from some later use on real-world data or a specific computing environment. Practical realities will influence choices around model size, speed, and architecture given constraints on deadlines, training time, and available benchmarks and computing resources. Using the engineering of the systems in support of a multidisciplinary research facility focused on the the estimation of human behavior, affect, and intent as a motivating example, in this talk I discuss techniques for the design, integration, and optimisation of suites of ML models for realtime inference in a production software system.

Bio:

Andrew Haigh is a R&D software engineer whose efforts are focused on the design and realisation of the software powering the National Facility for Human-Robot Interaction Research. He comes from an industry background in the development of low-latency and high-availability distributed systems in applications such as payment processing and automated trading.