[notsuoh]

Not an ML person, but I generally understand the concept. Why do you care about ceilings here? I get that the camera is pointed weird ways, but if you're trying to detect nudity presumably if trained on a bunch of images of that subject, it would just call ceilings False.

[sanxiyn]

No, because training set does not include ceilings.

[notsuoh]

I think I understand, so you basically need in your training set anything that the production model would presumably see for it to work well? You can't just say here are a bunch of positives and a bunch of negatives, the negatives have to be actual things the model will see.

[sanxiyn]

If you think about it, this can't be otherwise. Positive/negative is misleading, since that distinction is the entire thing we are trying to teach. Let's call two labels A and B. You tell the model:

This long penis is A. This short penis is A. This cat is B. This dog is B. Now, what is this ceiling?

The model, looking at the ceiling, discovers a long fluorescent tube. It is long! Neither cat nor dog is long (The model is yet to discover longcat), and while penis comes in long and short variety, all penises seem long-ish. The ceiling is A.

[jononor]

In the typical ML task formulation, closed-set classification, the model is forced to make a choice between the provided categories, even if the output is not really either. And neural networks tend to perform erratically outside of the training data distribution.

Adding common inputs to the training (or at least validation and test) sets is a good solution. Its hard data work, but will pay off. There are some techniques outside of closed-set classification that can help reduce the problems, or make the process of improving it more effective:

- Couple the classifier with a out-of-distribution (novelty/anomaly) detector. Samples that score high are considered "Unknown" and can be flagged for review. - Learn a distance metric for "nudity" instead of a classifier, potentially with unsupervised or self-supervised learning (no labels needed). This has higher chance of doing well on novel examples, but it still needs to be validated/monitored. - Use one-class classifier, trained only on positive samples of nudity. This has the disadvantage that novel nudity is very likely to be classified as "not nudity", which could be an issue.

[sanxiyn]

Yes, with emphasis on "well". It does work without ceilings. But your training set should include everything you will frequently encounter in test set. Ceilings are too frequent to miss.

[rubatuga]

If your model supports Out of distribution detection, you are very correct.