Why the first thing to do is not to get human-level performance?

In the week2 quiz, the 1st question, it says the first thing to do is not to get human-level performance. But why?

The human-level performance is the benchmark for measuring the avoidable bias. If there is no human-level performance, we can’t even get what the avoidable bias is after training a basic model. So I want to know why the first thing to do is not to get human-level performance?

Hey @Even,

Imagine we trained a model and got 99.99 accuracy (assuming a classification task). In such a case, we wouldn’t care much about avoidable bias. In a real-world scenario, time is limited and we usually know satisfying accuracy (for our use case).

3 Likes