Скачать или смотреть Your First ML model in Production Considerations & Examples

💻 Abstract:
As Data Science professionals, we want to do innovative, impactful work. Thus, our work on data munging and building machine learning models cannot happen in isolation from business objectives and the infrastructure of our organizations. In this talk, I will explore ways to identify impactful, executable Data Science work, and how to take this work to production. I will discuss what it means to have a model in production, including ways to score the model in real-time versus batch. I will discuss sample architectures required to make model scores available for your application, such as through an API or database. Finally, I will tie everything together with some of the processes and frameworks that allow for iteration and testing to complete the full life cycle of model deployment. I will provide a real example of taking an ML project all the way from data capture to real-time scoring in production.

🔊 Speaker bio:
Sabina Stanescu
7+ years working as a Data Science professional.
Currently manages a team of Data Scientists for consulting engagements with customers across use-cases such as Credit Risk, Marketing, and Engineering Process Design.
Previously worked at Points as Lead Data Scientist in Marketing and as Lead Product Manager for Machine Learning, integrating machine learning into Points' products.
MSc in Ecology, University of Guelph, with a focus on data analysis and modeling in ROI

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
0:36 Getting to know Sabina Stanescu
1:40 Taking the machine learning model and how you think about creating it.
3:15 Figuring out if it's the right time to build the machine learning model for your system.
4:48 Example on how to think about integrating your KPIs with model objectives.
10:27 Building and deploying that first pipeline.
11:26 Simple batch example.
13:37 Going from batch to real-time.
15:39 Things they need to think about, aside from the tooling are the business needs.
17:58 Prototyping and subject matter
18:34 Managing the pipeline with the Git lab pipeline and CI CD tools
24:20 How do you deal with engineering new features, launch and iterate, and add these new features to the pipeline.
28:57 A/B testing
29:54 Closing the feedback loop.

❓ Q&A Section ❓

32:20 What are your thoughts on ML run?
33:22 Have you used cube flow? Thoughts?
33:38 Were the KPIs tested on a preset or from recent history?
36:04 How do you make sure that you have similar sets for A and B and do you switch them back and forth?
37:31 Does the model get tested by the builder?
38:30 For A/B testing, do you also use a control group to measure incrementality?
41:25 Can you name a few tools that you would use today to run the same model in production?
42:59 How big is your model in production versus prototype?
44:15 How would you encourage your data science team to use traditional SDLC best practices?

45:43 Closing remarks