Frédéric Dubut of Bing said on Twitter that if you do use machine learning in production, like Bing does, you have to be okay with the results not being perfect. And not all applications are okay with not being perfect, but I assume search can be okay with it.
Frédéric Dubut wrote "To use ML in production you also need to be comfortable with a model that *will* get it wrong occasionally." "There are many applications where the required precision is just 100% - and unfortunately quite a few of these are still using ML," he added. These applications that have to be perfect are often in the financial, health and other areas. Think about the software that helps the airplane fly or software that helps you transfer money from on place to another.
Here are the tweets within the context:
To use ML in production you also need to be comfortable with a model that *will* get it wrong occasionally. There are many applications where the required precision is just 100% - and unfortunately quite a few of these are still using ML...
— Frédéric Dubut (@CoperniX) December 10, 2019
I was just thinking traditional software systems. A lot of real-world applications (e.g. banking) are better served with robust, well-tested and sometimes even provably-correct algorithms.
— Frédéric Dubut (@CoperniX) December 10, 2019
As you know, Bing uses a lot of machine learning in search - upwards of 90% or more.
Forum discussion at Twitter.
