Exponentially Weighted Averages, why set V0 as zero and then correct?

Course Q&A Deep Learning Specialization Improving Deep Neural Networks: Hyperparameter tun
1

I just went through the topic videos (exponentially weighted averages), and it seems to me that setting V0 to zero and then using bias correction (that requires some computation) to fix the results is weird.

Setting V0 to theta0 (the first reading itself) does a better job and doesn’t require any extra computation, just one simple initial value.

Since I know the people behind this course are very smart, I assume there is a good reason for that.

What am I missing?

3

Averages
Averages1285×1146 65.5 KB

This is a noisy cosine with an added positive random noise.

Bias correction cannot be done on V0 (cannot divide by zero), so it is still 0, and then applying this correction throughout the entire process is really unnecessary as it produces the same result.

See the green line where V0=Theta0, it adds no additional complexity and requires no additional calculation, and does a good job starting from V0.