How to prevent model from recognizing false Classes

Question

Let's say that I have a model that can recognize Cats and Dogs. However, when I use a picture of a Cup or Human it generates a random prediction at 0.70 confidence. Should I use sigmoid instead of softmax?

Or is there any fix for preventing massive confidence on new class object?

Answer 1

I haven't actually played around with this problem, but here are a few resources:

• AI.se question - Why not teach to a NN not only what is true, but also what is not true?

• How to Make a Neural Network say "Don't Know"? (maybe try requesting fulltext through ResearchGate )

• blog post - Making Your Neural Network Say “I Don’t Know” — Bayesian NNs using Pyro and PyTorch

Edit: I just found the following questions here on DS.se:

• Multi Class + Negative Class Image Classification Strategies

• Fine-tuning a CNN for recognizing two classes, but also being able to tell if none of them is present in an image