Week 2/Question 10

Hi all

I think there is a little risk of overfitting because we did not generate/simulate one fog pattern for each data. so it is probable that 1000 foggy sample downloaded from the internet, is only a sample of whole possible foggy states. then there is a risk of overfitting on those 1000 foggy states.

according to lecture, this is not detectable by human eye.

but when i select that option, it was not correct.

Hey @saiman,

Here is a thread with a similar discussion. You may find it interesting.

thanks a lot for quick replies :pray:

you mean that 1000 foggy images cover whole possible foggy states. in other words, not a tiny subset of the sets of all possible foggy states. is it correct?

To cover all the possible cases seems infeasible :slight_smile:

A reasonable assumption in the quiz that one would be able to download various foggy images that would be somewhat close to ones the model would see after deployment. The assumption is based on the fact that humans are quite good at visual perception.

I see…maybe I am trying to look at this concept from a too much theoretical point of view :slightly_smiling_face:.
but anyway the same limitation of hearing system (limited frequency) applies to visual system also(we are unable to see NIR frequencies and a lot others).

so what I understand is the key is that our sample should be something near (as much as possible) to what our deployed system will see/hear in the future.

just for the sake of discussion and as an example suppose in the rear-mirror example if we could guarantee (I am aware that we can not even all 10000 hours of noise car sounds similar to our ears) that one hour is a good representative of whole (real world) possible car noises, then it would be free of overfitting risk.

thanks a lot for answering kindly.

True. We would also consider the noise acceptable in this case :slight_smile:

1 Like