Question 9/27 fast.ai v4 lecture 5
Why is it important to understand the issues around the use of metrics to track performance?
Because much of AI/ML centers on optimizing a metric. Understanding the issues that can arise when pursuing improvement of a metric is fundamental to rolling out AI/ML systems. The more we know about this area, the higher the chance our system will not act in a completely surprising (and ofen counterproductive way) and if it does, that we will be able to pick up on this behavior and implement countermeasures.
Relevant part of lecture
The problem with metrics is a big problem for AI
- a blog post by Rachel Thomas Reliance on Metrics is a Fundamental Challenge for AI
- a paper by Rachel Thomas and David Uminsky Goodhart's law
- relying on a metric makes it less reliable; when a measure becomes a metric, it ceaseto be a good measure. How Algorithms Can Learn to Discredit the Media
- a blog post by YouTube's ex-engineer on unintended consequences of optimizing a metric