How to detect that a neural network input is out of scope?

Question

I have built a neural network that detects dog breeds.

My issue is that I don't know how to handle images that don't contain dogs. Indeed, whenever I input a flower image, it will be classified as a dog breed.

Is there a way I could train my neural network to the output "I don't know"?

Attempt:

I have tried adding pictures from other objects than dogs. Unfortunately, I don't know what proportion of external images with the not_a_dog label I should add to correctly detect non-dog images. I don't want to create a bias in my data.

Answer 1

Rule of thumb is to create a balanced dataset with both positive & negative samples. There's no harm in using the same number of dog images(positive samples) & images without a dog (negative samples).

Unless you add negative sample to your training data, there's a very high chance that your model will predict "dog" for unseen data without a dog.

You can use probability as a measure to avoid negative samples. Usually if you have a good training data, the predicted probability of "dog" should be very less for negative samples.