Question 30/30 fast.ai v3 lecture 10

In the context of exponential moving average, what debiasing or bias correction mean?

Answer

The initial mean is initialized to 0, which has a big impact especially for the first couple of samples. We 'debias' the mean to correct for this.

Relevant part of lecture

supplementary material

There is a nice video from Andrew Ng (the deeplearning.ai series) on this. Here is a screenshot taken from it:

The red line is a regular average, and purple and green lines are EMA without correction and with correction. The impact of the inital mean of 0 is clearly visible if you compare these two close to the origin.