random seed for regression NN

[rneeser]

Hi there,

I have two questions regarding the (global) random seed:

• Why do we have to set the random seed before creating etc. our models in part 2? I don't see where we have randomness there.
• Why do we need to specify the global random seed again and again? Shouldn't it work for the whole notebook when everything is run and active?

Thanks for your help!

[mrdbourke]

Hey there, great questions, I'll answer both.

Why do we have to set the random seed before creating etc. our models in part 2? I don't see where we have randomness there.

When models get created, they get randomly initialized with different weights and biases (I often refer to these as patterns but their proper terms are weights and biases).

What this means is:

model starts with random numbers -> looks at data -> adjusts random numbers -> looks at data again -> adjusts -> looks -> adjusts...

And it keeps repeating that cycle until it finds ideal patterns (weights and biases) to represent the data it's looking at.

The random seed for creating a model means that the model starts with the same random numbers to begin with and then adjusts them in the same fashion whilst it looks at the data each time.

This is not 100% necessary but it makes sure the numbers I get on the screen are as close as possible to the numbers on your screen.

Otherwise, because of the inherent randomness of machine learning, we'd get slightly different results each time.

Why do we need to specify the global random seed again and again? Shouldn't it work for the whole notebook when everything is run and active?

Think of each notebook cell as its own program. Each time you run a cell, it runs and then it's done. The variables stick around but the seed doesn't.

That's why we need to set the random seed each time we initialize something new.

Let me know if anything needs clarifying.

[rneeser]

Thanks for the quick reply Daniel! That makes sense.

So, when you work on a project you would also set the global random seed so that your experiments for improving a model are comparable? Both between different experiments (and initializations) as well as with your colleagues.

Thanks for the clarification!

[mrdbourke]

So, when you work on a project you would also set the global random seed so that your experiments for improving a model are comparable? Both between different experiments (and initializations) as well as with your colleagues.

Yes that's correct, that way if you shared your results with someone else, they'd know how they can replicate them