Lecture from an MLOps course (Week 2, Lesson 1) introducing best practices for making machine-learning models production-ready. Covers model selection and training, error analysis, handling skewed datasets, and contrasts model-centric vs data-centric development (including an anecdote referencing Adam Coates). The instructor emphasizes using data-quality improvements (targeted, efficient edits) alongside model changes to reach deployment goals.
Chapters
In this week you learn about some best practices for building a machine learning model that is worthy of a production deployment. One of my friends, Adam Coates, joked that the way he listened to me give advice to machine learning teams, he felt the way I gave advice was quite consistent from project to project, so that he could almost replace me with an if then else sequence of statements. And I found too, when several senior machine learning engineers look at the project, the advice they tend to give is also remarkably consistent. What you learn in this week is what are some of the key challenges of trying to build a production ready machine learning model? Things like how do you handle skewed data sets or what if you do well on the test set, but for some reason that still isn't good enough for your actual application. And I hope that after this week's materials you'll be able to very efficiently know how to improve your machine learning model to solve the most important problems that then make it deployment ready. Let's dive in. This week our focus will be on the modeling part of the full cycle of a machine learning project and you learn some suggestions for how to select and train a model and how to perform error analysis and use that to drive model improvements. One of the themes you hear me refer to multiple times is is model centric AI development versus data centric AI development. The way that AI has grown up, there's been a lot of emphasis on how to choose the right model, such as maybe how to choose the right neural network architecture. I found that for practical projects it can be even more useful to take a more data centric approach where you focus not just on improving the the neural network architecture, but on making sure you are feeding your algorithm high quality data and that ultimately lets you be more efficient in getting your system to perform well. But the way I engage in data centric AI development is not to just go out and try to collect more and more and more data, which can be very time consuming, but to instead use tools to help me improve the data in the most efficient possible way. You learn some ways for how to do that in this week. So I'm excited to go through this week's materials review on training models, but first let's look at some key challenges that many teams face when building machine learning models. By understanding these key challenges, you'll be better able to spot them ahead of time and adjust them more efficiently for your projects. Let's go on to the next video.
