为什么大多数机器学习项目无法投入生产环境以及如何克服这些困难.pdf

1、Why Most Machine Learning Projects FailandHow to Beat the OddsWenjie Zi,Senior machine learning engineer|Grammarly12024.11.1810 years of industry experience as an Applied Scientist and Machine Learning Engineer.Senior machine learning engineer GrammarlyDeep learning instructor University of Toronto(

2、certificate program)Co-founder of Toronto AI Practitioners NetworkWenjie ZiAbout meOpinions Are My OwnQ:Have you worked on any machine learning projects?Q:Have you worked on any machine learning projects that didnt make it to production?Some projects end up delivering significant value;Many others d

3、o not.My journey developing machine learning projects that drive social media platforms,fintech solutions,and productivity tools2.1 2023 Rexer Analytics Data Science SurveyFigure:proportion of machine learning models deployed to production,as reported in 1.Failure rate of machine learning projects2.

4、Fail fast:based on the experimental results,there werent sufficient positive signals,leading to the project being quickly discontinued.But some failures should be celebrated!2.What leads to major failures in machine learning projects?And how can we prevent them?Agenda1.Overview of machine learning p

5、roject lifecycle2.Five common pitfalls and how to improve:Tackling the wrong problemChallenges arising from dataStruggle to turn model to a productOffline success,online failureUnseen non-technical obstacles3.SummaryAn overview of machine learning project lifecycleFigure:a simplified high-level diag

6、ram illustrating the machine learning project lifecycle.Machine learning project lifecycle2.The lengthy multi-step process involves many handovers,increasing risks due to complexity.This data-centric optimization requires feedback signals to guide iterative improvements.Five common pitfalls and how