Why do we use Conv2D 16, 32, 64 filters and Dense 128, 256, 512 neurons?

I​s there any reason that number (16, 32, 64 filters) (128, 256, 512 neurons)?

W​hy don’t we use Conv2D 15, 30, 45 filters or Dense 100, 200, 300 neurons?

A​nd how do we decide to use the numbers of filters and neurons?

T​hank you

1 Like


There is no rule of thumb on deciding the number of filters like 16, 32, 64… it could as well be like 15,30,45 etc. It is generally a good practice to increment it by two fold. No. of filters or no. of neurons are the hyperparameters which you could choose by yourself. Generally it is emperical meaning you go with a number on basis of your prior experience and then tune it to reach your expecred performance.