Open Office Hour: Q&A on Machine Learning
About this course
An hour with a senior trainer to answer your machine learning questions. There is no fixed agenda, because the session is whatever the people in it need. Bring a model that is misbehaving, ask which approach suits your problem, or listen while we work through someone else's question.
Questions range from beginner to advanced. The mix is part of the value. Why a model looks too accurate is worth hearing about, whether or not it is your model. The most common questions are about results people do not trust. An untrustworthy result usually comes from a validation problem rather than a modelling one, and is faster to diagnose out loud than over email.
You can also arrive with nothing and listen.
The session is held online, so you can join from wherever you work.
Who this course is for
Anyone working on a machine learning problem in Python who wants to put a question to someone who does this for a living. Questions run from first-model beginner to production-scale, and both are welcome in the same hour.
This session also suits teams with a model that is not behaving and no in-house specialist to ask. Examples are a metric that looks too good, a model that degraded after deployment, or a choice between two approaches that nobody can settle.
Prerequisites
None. You do not need to have built a model before, and you do not need to have taken a course with us.
You will get more out of the hour with a specific question or problem in mind, such as a model that will not converge, a result you do not trust, or a decision between two approaches. Listening in is also fine.
What you will learn
- Get a direct answer to a machine learning question from a working practitioner.
- Work out why a model is behaving in a way you did not expect.
- Judge whether a result is real or an artefact of how the data was split.
- Decide which approach or library suits your problem, and what to learn next.
Syllabus
Day 1: One hour, driven by the questions in the room
- Questions taken in turn, with time given to each rather than a fixed agenda
- Diagnosing a model that is not working, or a result that looks wrong
- Choosing between approaches, libraries and frameworks for a given problem
- Validation, leakage and the results that are too good to be true
- What to learn next, and what to skip for now