Executive Briefing: Why machine-learned models crash and burn in production and what to do about it

Much progress has been made over the past decade on process and tooling for managing large-scale, multitier cloud apps and APIs, but there is far less common knowledge on best practices for managing machine-learned models (classifiers, forecasters, etc.), especially beyond the modeling, optimization, and deployment process once these models are in production.

A key mindset shift required to address these issues is understanding that model development is different than software development in fundamental ways. David Talby (Pacific AI) shares real-world case studies showing why this is true and explains what you can do about it, covering key best practices that executives, solution architects, and delivery teams must take into account when committing to successfully deliver and operate data science-intensive systems in the real world.

Topics include:

• Concept drift (Machine-learned models begin degrading as soon as they’re deployed and must adapt to a changing environment.)
• Locality and limited reuse and generalization of models
• A/B testing challenges, which make it very hard in practice to know which model will perform better in production
• Semisupervised and adversarial learning scenarios, which require modeling and optimizing models only once they’re in production
• The impact of all of the above on product planning, staffing, and client expectation management

This session was recorded at the 2019 O'Reilly Strata Data Conference in San Francisco.