Hamza Tahir - Why ML in production is STILL broken? скачать в хорошем качестве

Hamza Tahir - Why ML in production is STILL broken?

💻 Abstract:   Why ML in production is STILL broken? Around 87% of machine learning projects do not survive to make it to production. There is a disconnect between machine learning being done in Jupyter notebooks on local machines and actually being served to end-users to provide some actual value. The oft-quoted Hidden Technical Debt paper, by Scully et. al., has been in circulation since 2017, yet still, ML in production has ways to go to catch up to the quality standards attained by more conventional software development. This talk will aim to break down the key aspects of what sets machine learning apart from traditional software engineering, and how treating data as a first-class citizen is a fundamental shift in our understanding of complex production ML systems. 🔊 Speaker bio: CTO, maiot GmbH Hamza Tahir is a software engineer cum machine learning engineer based in Munich, Germany. He has a passion for trying to connect the dots between his various learning experiences and to continually learn and grow from new challenges. Hamza is currently co-founding his ML startup, maiot, with the aim of bringing proper software engineering practices into machine learning workflows. If you enjoyed this talk, visit us at https://mlopsworld.com/ and come participate in our next gathering! 💼 Would you like to receive email summaries of these talks? Join our newsletter FREE here: http://bit.ly/MLOps_Summaries 📧 Timestamps: 0:00 Intro 0:23 Getting to know Hamza Tahir 2:35 The hidden technical debt paper 3:38 Why 87% of ML projects never making it into production 5:05 Traditional solutions for technical debt 5:55 Debt in ML is more Complicated 8:54 The generic ML journey 9:08 The Baseline 11:00 Data Changes 12:44 Model Goes Stale 14:28 Upstream Feature Changes 18:17 Systems tend to fail faster, harder, and silently 19:28 SOTA Production-ready ML architecture 21:19 ML in production is (still) broken ❓ Q&A section ❓ 22:17 Data structure and also meaning 24:27 Can you go into more detail about the framework and solution to resolve these issues? 27:25 Where the 87% code is coming from. 28:22 How can I learn more about the feature store? 29:19 How do you view two open-source tools like ml flow? 30:35 Is it better to have an opinionated architecture? 33:00 Considering the fact that you're in Munich, have you considered adding additional or possibly multiple boxes, which are titled beer in the architecture, in order to ease the transitions between various boxes in dealing with ml projects? 33:33 Do you have suggestions on communicating with stakeholders? 35:08 What are your thoughts on handling data drift? 36:48 Any advice to avoid having to rewrite pre-processing? 39:01 What is the cost associated with implementing this ETL pipeline? Rough order of magnitude? 40:11 May you explain the distributed learning in a product? How we can manage data, and the learning process easily? 42:30 Closing remarks